Anthropic S-1 IPO, OpenAI training security, AWS capacity sold out

This transcript was generated automatically and has not been fully reviewed. Timestamps are approximate and wording may contain errors.

00:07 Welcome to The Cloud Pod, where the forecast is always cloudy. We talk weekly about all things AWS, GCP, and Azure. We are your hosts, Justin, Jonathan, Ryan, and Matt.

00:18 Episode 378, recorded for October 27th, 2026. Agent Compute Engine gives your AI a padded cell. Good evening, Ryan.

00:29 Good evening. I feel like I need a padded cell.

00:31 I definitely need a padded cell, but that's why we brought Josh here to secure us.

00:36 Hello, Josh. You supervise, right?

00:38 Uh, that's a terrible life choice.

00:40 We always say that, you know, the inmates are in charge of the asylum when Justin's gone. So like, no, this is, it's fitting.

00:48 Josh, do you want to say hello and introduce yourself?

00:50 Hello, I'm Josh. So Yeah, I'm happy to join you guys. Known you for a little while and I've been working on cloud for a little over a decade. I've had some fun working on the edge for a large multinational restaurant, getting some Kubernetes on the edge, which was a blast. I got to work in the startup life for a little while, then, you know, lived in the SaaS world for a little bit. Now I'm trying to help secure the world with all of the fun we've had lately. Last week was a blast.

01:16 I don't know, you saw the .NET, which I don't even think is in our show, the .NET vulnerability today and There was what, PHP, and there was a Node.js one last week, so you must be having fun.

01:28 Yeah, yeah, no shortage of things to do. And there are days where I'm like, I'm kind of glad I'm not quite on the front line so much, but I feel for the folks out there.

01:40 All right, well, let's kick us off then with follow-up news. Speaking of security and securing everything, OpenAI actually came out with an article about securing our frontier training models. And if you remember from the last, I don't know, last week, 4 months ago, 5 months ago, when every model has, you know, taken or tried to take over the world while they were training it, you know, and trying to benchmark it. OpenAI actually published, I guess they're listening to the show, Ryan, after last week that I don't think we dropped it yet, but when I said,

02:13 oh, we've been asking for this for weeks.

02:15 That's true. But we said, look, they really need to step up their control. So OpenAI listened. Apparently we have that much influence. So congratulations, Ryan, the world is screwed.

02:26 Naturally,

02:27 about how they are hardening a Frontier model training environment after the Hugging Face incident, including red teaming campaigns that use their own strongest model against its own infrastructure. Policy change also occurred. Frontier models training now runs with no live internet access, which I thought they did before, and will stay until alignment research, monitoring, and internet infrastructure security justifications broader requires broader access. Follow-on from the incident on October 20th, an internal research model exploited a DNS resolver misconfiguration again— well, I guess not again, but another incident— to gain limited live internet access during a training run. OpenAI contained it, paused the training run and evaluation of its most capable model, and resumed reinforced learning Afterwards, red team results came with agent red teaming, including with the upcoming GPT-6 Azure Cyber model. Were given source code and configuration details, spent hundreds of billions of tokens, just a normal day in the world of OpenAI, probing the infrastructure. They found previously unknown vulnerabilities, misconfigurations enabling micro VM escape, routes to the internet through metrics, logging, and authentication services, and cross-sandbox communication via CPU cache timing, which is— holy crap, that's insane. Intense.

04:01 It's bonkers.

04:03 Yeah, there's more. You should read it in the show notes. It goes into more details, but I read a little bit about the timing one. But Ryan, I know you did some additional digging on this one. Why don't you fill us in?

04:14 Yeah, I mean, this is something, like we said, we've been asking for for a while. And, you know, like, it's— it is weird that they're saying, oh yeah, no, we're they're removing the internet access when pretty sure they've said that a lot. But, uh, they get into a little bit more detail into how they're removing internet access, which is pretty fascinating because for their, at least for one aspect of their reinforcement learning process, which I guess is where these things, uh, were probably the most exploited, they are basically creating an offline version of the web and they're calling it an offline world. And so they've recreated web indexes and they're caching web content and basically, I hope, maybe it's a smaller version, but they're basically using a digital twin of the internet, which is crazy. And so that way they can sort of have it completely isolated, but still be able to do things like web queries and searches and provide that sort of level of detail, which is sort of nuts.

05:17 So they're putting the simulation in a simulation.

05:20 Correct.

05:21 So they're building the matrix.

05:22 You never needed more proof that we're in a simulation. It's turtles all the way down.

05:29 Yeah, it is interesting. Just like any foothold these things can get, they seem to just exploit something and then just bounce their way out. It's—

05:37 I mean, they found a CPU cache timing bug. Like, that's an insane detail to figure out. And then to do, I mean, I half wonder how much, you know, they say billions of tokens, but like how many tokens do they actually use? And then a lot of the vulnerabilities out there, you know, more at a macro level is, you know, are these things actually exploitable? Like,

05:59 oh, I mean, that's the, that's the biggest issue by themselves. No. Right. But that's the power of AI is they're chaining them together to create these exploits.

06:08 Right. It's the chain level attacks. You know, you take one little thing and everything goes bonkers from there.

06:13 Yeah. And so my day job, it's awful now 'cause it's like, well, we used to just basically ignore, you know, like the stuff that was, you know, not reachable or low or medium and, and, and that. And now you have to take a whole new look at your risk profile and be like, well, if, if these 6 lows can be easily exploited by AI, are they still low, right? In context. And how do you, how do you measure that context? And it's tricky. So it's, there's a lot more of those, you know, lower severity CVEs that can be chained together. And I don't think I'm alone in sort of deprioritizing those and—

06:48 No.

06:49 To make sure that we can get the highs and the criticals patched. So it's this like, uh-oh.

06:53 I used to argue with our, you know, different security teams being like, look, I fixed all the highs. Like, we're trying to fix the root process here and then we'll get into the lows. You know, once we fix the root process, we have a, about a proper bakery, Docker, ECS, CI/CD pipeline, you know, but like now you can't wait and you're back to like, oh my God, you know, it's like back in the day of EC2 Classic, you know, if anybody else remembers those days, Ryan, I think might here, but everything was public. So you had to care about all those. And then the VPC at least obfuscated some of that, you know, behind private subnets and things like that. Thank you Azure for releasing that last year, you know, but having private subnets be able to actually, to handle some more of those things.

07:34 So yeah, well, I just, you know, it's, it's funny because you go into the cloud and you realize that your security model can no longer be this high wall garden, but then the first thing you want in the cloud is a high wall garden. Turns out your garden's a mess.

07:47 Don't look in my house. Okay. I will say, I think the most patched thing in my home network is my containers. I have them set every week. To just auto-update to the latest version. And the only problem every now and then is my Pi-hole server because it requires DNS. If it, the timing is too long when it reloads the container downstairs, it just kills my DNS in my house. I have to go ask SageMaker.

08:14 The whole internet dies in your house. Yeah.

08:16 And there's been multiple times I've been like, okay, my internet is down. Let me go reboot. Maybe UniFi had a problem or something. No, no, no. It's always Pi-hole. It's always DNS, guys.

08:26 Your Pi-hole has max 1 to 2 years to live before your kids are fully screen time and they're, and then that's going right out the window. You'll have, you'll be IT with like screaming users. Like, mm-mm.

08:40 There's a reason I moved out of IT. So, you know, I still get it from, I don't know, parents, wife, everyone still. So it's fine.

08:49 We do set up an at-home help desk. Go put a ticket in.

08:53 Ryan, how would your wife appreciate that? Hmm? Just when she comes up to you, you're probably, keep her ticketed for that. I'll get to it later.

09:02 Uh, well, I wouldn't be here right now. I think it'd just be an empty, empty place behind me. 'Cause I would be probably dead and buried in the backyard.

09:11 Well, speaking of empty places, after, I actually don't know how to say this certificate name company properly. Immutra.

09:20 That's what I've, that's what I've been saying. I don't know if that's right. Yeah.

09:24 After IMUDRA lost 4 of their private keys last week, they have officially dropped out of the TLS business altogether. They are no longer issuing certificates. The 8 compromised intermediary certs use P384 keys that should have carried 348 bits of entropy, but in actuality, they range from 3 to 8 bits. Minorly different, which was crackable with, I don't know, my cell phone in about milliseconds, maybe a few seconds on my cell phone.

09:52 It's the same password as on my luggage.

09:55 Remind me to change my luggage code. Anyone who, anyone who could recover keys could mint certificates for any website or mail servers that browsers or many other clients would just trust, which breaks the core guarantee of the TLS model. So I think this is the first time in a long time that A complete company has dropped out of the TLS game.

10:18 I've never heard of this before, but I think it's the right move. I mean, it's the, the passwords that were cracked, the level, like, is, is pretty embarrassing. And I don't know how they would, could be a trusted CA after that. Right. So it's like, makes sense.

10:35 I mean, it's not even that, you know, with certificates dropping down in length to what? I think what this year is like 100 and some odd and eventually like 47 days or something like that. I don't remember the numbers. You know, you have to have a full system. So at that point, if they already haven't invested in a full, I'm gonna call it Let's Encrypt style system, even though there's a better term for it, you know, they need to figure out, what was it? Acme. Acme, thank you. But whenever I see that, I just see Wile E. Coyote. Like, that's where my head goes. Natural. And I'm like, wait, is that wrong? You know, with Active Directory SAML certificates, you know, where if they haven't invested that yet, they're behind the game and there's no point doing it, honestly, at this point. So, you know, it doesn't surprise me they dropped out, especially after that, and they've lost all trust. But the problem is, you know, having these certificates, you know, Chrome required updates, you know, on every computer. Like, it's a massive— it's not just like they revoke it, it's a lot more stuff that had to be done too.

11:37 And it's been a real pain for like, if you have a third-party one that doesn't use the OS certificate store, now you got to go try to track all those down.

11:46 Oh, it is awful.

11:48 It's like vulnerability management. It's not just, you know, in your application code. It's not just the main one. Then you got to dig through everything to figure out what you can update. RubyGem hell. Sorry.

11:58 Except for that as well. It's like, you know, it's Node. Compiles certs and then a lot of time it's, you know, anybody you're fooling around with, you know, you're catching what they got.

12:10 So follow-up from prior shows, episode 361, does GitHub's new Immutable Subject Claim for Actions, OIDC, actually make things more secure? So the specific item that we want to follow up on, and Ryan has done the homework on-ish, determine that— determine whether a new— the new GitHub OIDC provider actually improves security or not. Ryan, tell us what you found.

12:37 So I don't remember saying this on their original show, so I don't remember what I was thinking, but, um, it's either I didn't fully understand what it was or I just didn't know why it was more secure, and those could— or both. And this is actually cool, um, because it is sort of like the, the previous sort of OIDC assertion validation was stuff like just repo name and stuff that's really easily guessable or fungible or, you know, that kind of thing. And this is actually, they've changed it to those claims when it's doing that validation and the exchange is much more strong where it's using, it's immutable and it's tied to your user identity, which is kind of interesting, like ownership repo identity rather. So that's, it's sort of interesting and it, It's basically going from like, you know, a plaintext name to a specific ID. So you can't recreate a repo with the same name as something that has a valid token and do the exchange. So it's cool. It is an improvement and I like it.

13:36 That's a pretty good improvement actually, because I never thought about like deleting a repo, recreating it, and then having, you know, immediately GitHub has access to that without really thinking about it. So

13:47 yeah, nothing you can, if you know, like it's a lot of knowledge that you have to know to actually chain that exploit together, but—

13:53 AI.

13:53 You know, if you can get a, exactly, if you can get a token issued because you're matching the claim, then whatever that token has access to, you have access to. So.

14:03 On to general news, Microsoft this week has issued voluntary buyouts for 60,000 people. If you've seen anything, I feel like this has been in mainstream news too. Microsoft announced, voluntary exits of roughly 60,000 employees, which tells cloud watchers a lot about the cost of AI and buildout. The offer covers eligible employees in the US, Canada, India, and the UK, experience level 70 and below. So if you're at Microsoft, hopefully that makes sense to you. About 27% of the Microsoft workforce, 200, which is 223,000 workers. Decision are due in early December, which departs on December 31st. This is the third workforce action this year. It's a very nice way to say that following voluntary retirement programs in April, over 30% of employees took, and in July when they had 4,800 layoffs. The new offer extends to the vice president level versus the senior director and below. That's where that 70 level means. I really should talk to my friends that work at Microsoft to see if this makes more sense than, you know, it's like Amazon's L3, L5, L7, you know, people are like, I'm an L7. I'm like, cool, do you want a badge? I guess you have a badge. Like, I don't really know what you want me to do with this. I was, I just remember talking to somebody at Re:Invent. They're like, I'm an L8. And I'm like, okay. And apparently that's actually pretty senior. Uh, I made up the number, but yeah. Anyway, back to this story that we were— before Matt went on a tangent— Microsoft is shipping its usage pricing of AI, saying that internal AI tooling for programmers, sellers, and supply chain workers is producing noticeable results.

15:53 I mean, it's kind of crazy to have a layoff without laying off, right? Like, that's— I think this is better than just laying off and cutting costs. I think it's allowing people to sort of self-select, which if you're close to retirement age or, or if you know, you're thinking about getting another job anyway, take an exit package. That's, that's pretty sweet. But it is sort of crazy that these large companies are making this level of move. It's a, a real big bet on AI and, and so they're funding all this infrastructure that doesn't quite make money yet. And it is nuts to think that this is what they're, You know, they're getting rid of the human capital, which is the most expensive part of running any business to fund it, which is kind of nuts.

16:37 It'll be interesting to see, you know, they say they don't have a target number, if there's going to be more action that follows after, you know, like, hey, I don't think this will be the end of it.

16:49 Definitely not. No, this is, this, if they don't get enough people that voluntarily self-select, then there'll be an actual layoff, right? Like, that's how that'll go down.

16:58 That'll be for Q1.

17:00 Yeah, I assume it's just like an optics thing. It's like, we didn't do layoffs, you know, but we gave people the option to leave. And I wonder if, you know, they say like, you know, the return to work, you know, where you force everyone to the office and whatnot, you're losing people that can go get a new job elsewhere that, you know, have the ability to go, and you're probably not retaining your best talent is one of the arguments against return to work. And Here I kind of wonder if this is trying to do the same thing, almost like trying to avoid that type of a layoff where you are, you know, just saying, yeah, maybe, you know, this is the second time you got this, so clearly you're on the hit list somewhere. Time to go find a job, take the money and go run, you know. And then it becomes not that, but I also just assume it's an optics problem here. Like, we don't want to say that we did two layoffs in a year. That kind of freaks stockholders out.

17:50 'Ah, you're so cynical. I think they're doing the right thing.' And I'm so much more of an honest intent, like, 'Oh, we'll give people the option before we do a layoff.' And it sounds like it's universally applied by level rather than sort of like you being an individual being singled out for like, 'Hey, you should maybe think about quitting.' Oh, I

18:09 didn't realize that. So you think it's just like broad spectrum, like if you're X level and below?

18:14 Oh, 100%. Yeah. That's a— and that's It's, uh, that's what the first one was as well. It was just a different level.

18:20 Yeah. Okay. It's a crazy amount of people though.

18:24 Like, yeah. Like

18:26 it's like one third of their global workforce, I think. Or was it like one third of the US? Yeah.

18:30 27% of their global workforce. Yeah. Like that's, it's crazy to think that 60,000 is a third and then just like, that's a big number to downsize.

18:41 Mm-hmm.

18:43 And I wonder the distribution of that, those levels, like it's, you know, that for the everyone else that's not taking that package, right? Absorbing that extra work or doing what else is, is interesting. And so if there's concentrations of people accepting those packages, it'll have some impact.

18:57 But AI can do it for you, Ryan.

18:58 Didn't you know that? I don't do any real work anymore.

19:02 I wonder if they got kind of the other side of that. Like if 3 people on your team leave, like you're not getting that, like we're resetting your offer. We're not letting the whole team go.

19:11 Oh, I never thought about that. Like, the entire department leaves for something and they're just like, oh no.

19:17 Yeah, yeah. Just like, literally they're like, let's do it, we've always thought about like starting a petting zoo, and just like the whole team leaves.

19:24 Well, that's when you get into retention bonuses. So it's like, I think that's good.

19:29 Yeah. So I'm just kind of curious if they've got anything with that in mind.

19:33 I'm sure somebody has thought through it, you know, their HR department is probably more people than, you know, the companies I've worked for in the past.

19:42 So, you know, right. Yeah. The HR department is bigger than my current company.

19:46 Yeah. Right. Like, I'm like, yeah, I hope, at least I hope somebody has thought about it. Yeah. But I also wonder then how many people are just moving over to be like contractors for them too. So like, are they ending up just moving it off one line item to the other? You know, we've all gotten the V- emails, you know, versus the whatever the actual employee email addresses and just defines whether they're, you know, a vendor or internal.

20:11 I mean, if I could get a package for then just getting a contractor label and doing my same job, that'd be pretty awesome. I'd do that in a heartbeat.

20:20 If you ever work for DOD, a ton of people work for the US government and then literally go take their same job back as a contractor the next day. I mean,

20:29 yeah, but they don't typically give you a retirement package.

20:32 Yeah, well, that's true.

20:34 Yeah, that's the part I like. I don't care about becoming a contractor, but give me a, you know, a 3 to 6 month severance and I become a contractor, I'm in. All right, moving on to AI is how ML makes money. So Databricks just made AI Runtime generally available, bringing fully managed GPU training infrastructure directly into the lakehouse. And this matters for any team tired of running a separate GPU stack next to their data. AIR, or AIR, not really sure, is generally available and it's a fully managed GPU service where Databricks handles all the provisioning, node health, and workload optimization. You can access it via CLI, notebooks, IDEs, and it supports Ray and bring your own container. You can even spin up a multi-node RDMA cluster with dozens of GPUs in seconds. Reported customers results in reported customers include Superhuman moving from 8 A100s to one node to 16 H100s across nodes, whatever that means. And it cut training for that from them from 7 days to 48 hours, which is pretty sweet. Governance is the main differentiator. The URD catalog enforces permissions on data checkpoints and logs for GPU workloads and the training space where data lives with no extra copies. Isolation is at the VM level rather than I showed shared host kernel and PrivateLink and serverless egress controls. Post preference and OpenAI Hugging Face incident are being addressed. The developer workflow features include a native Lakeflow orchestration, which means you can do just a few lines of YAML to productize, which is, you know, piles of YAML all replaced, great. It includes a new MLflow multi-node log viewer with a watch flag and a GenieCode agent skills for debugging and performance tuning because no one does any real work no more. The cost visibility comes through Databricks billing system tables and the workload level usage records. Named adopters include Robinhood, PicBay, AmeriFlow Health, TinyFish, and other big names that are name dropping to be like, look at our cool new feature. Good job, Databricks.

22:40 I will say the best naming convention that anybody's come up with is GenieCode. I'm just drinking— every time I see it, I'm like, I Dream of Jeannie. Like, that's what pops into my head, and it's just developing for me. Like, sorry.

22:55 Yeah, makes sense.

22:57 There's a real conversation here, but like, I just think of I Dream of Jeannie and Nick at Nite. Like, that's where my brain goes.

23:03 I mean, it makes sense. Yeah.

23:06 I mean, look, the performance people are getting, you know, I also think you're just getting that as you move You know, for, I mean, I believe the A versus H100s is just a newer, you know, GPU and things along those lines. You know, I think it's great you're getting better feedback. I mean, I complain that it takes me 3 minutes to boot up a Spot instance and I'm yelling at my computer, why are you taking so long? So I can't imagine 7 days of training to find out you trained your data wrong would drive me absolutely crazy. I mean, 48 hours is still more than my attention span is. But it's nice to see that they're really making these improvements here on the platform.

23:43 Yeah, I wonder if a lot of this is coming from like with DRA came out recently for like Kubernetes to, you know, really improve the utilization of, of these chips.

23:53 Oh, it's definitely tied to Kubernetes under the hood, right? Because you're seeing the same type of enhancement on a lot of these different platforms, which just lets you know that they are you know, their platforms are hosted as on top of Kubernetes for management.

24:08 Yeah. They just didn't realize for the longest time, like how much waste there was on all these chips, just because of the way the training worked. There are just so many sitting idle or not being used. Seeing these be used a lot more efficiently is going to be great.

24:23 Better power consumption, better performance. Can't complain.

24:29 Introducing Claude Duet. Anthropic just opened their research preview. It's a proactive agent that works in the background on your behalf, and it's worth a look for anyone tracking agentic AI in the enterprise and the small business space. Claude Duet is a standalone iOS and Android app and Claude AI experience built on Claude Opus 5.5. It sends a morning brief, it monitors connected apps every few hours, and messages you only when something needs your attention. Waitlist access starts with US Pro and Mac subscribers, and it shares the existing usage limits for your account. It runs tasks on its own cloud-hosted virtual machine, dedicated browser, so work continues when your, when your phone or laptop is off, and tasks range from scheduling to building dashboards, forecasts, web apps, etc. Integrations use the 6,000+ connectors that are available in the Cloud Marketplace, and the safety design is a key discussion area. Where credentials follow managed agent models, so the agent never sees passwords or tokens, and an auto mode classifier gates all actions that it takes. A new dedicated classifier blocks sharing of top-tier sensitive data like credit card and government ID numbers, with red teaming and external pen testing and bug bounty coverage are included, which doesn't make any sense. I'm hoping that that means they just validated and performed red teaming on their product instead of like, can make my AI agent red team Anthropic's, you know, ecosystem.

25:52 I'm looking forward to testing this. Thank you, Bolt, for now telling me I have to actually go talk about this at one point in the future. Um, but I mean, I signed up for the beta. The piece of it that does bother me is that it's running on Opus 5.5. It's an expensive model to— it like, it is consuming your, your main limits. So it just feels like it's going to be fairly heavy. So like, you're gonna wanna be careful of how many things you connect it to when you do set it up. 'Cause I've seen people that have had access, like just reading online about it, and people are like, oh yeah, I've like started my day, I'm already 20% into my weekly limit and I haven't done anything yet. And it's like, oh, well this is running and doing things and whatnot for you on the backend. And they're like, great.

26:37 Yeah.

26:37 So Matt's gonna wake up to no tokens. Like, he's already done by like noon most days, so like

26:44 only the 5-hour session, Lambda, it's fine.

26:46 Yeah. I didn't, yeah, it, I, this seems to be more of a, you know, a hot trend, right? This is Anthropic's the last to do this. You know, this OpenAI has Dots, Facebook did Muse. And so like, this is definitely, I think, where a lot of people that are not technical or really want to use AI, which is like, you know, help manage my life. And not excluding technical users, 'cause you know, I think, you know, if anything like me, then I need this more than anyone else. But I, you know, I still don't know if I trust, you know, sending all my data passwords and stuff into, you know, the, some other ecosystem. But you know, I think the benefits might outweigh the cons, but it's definitely something I need to think heavy on and see what I can do with it.

27:32 I mean, I've definitely started limiting more of what I put into SaaS products over the, like the last couple years and With AI, you know, part of the reason why Josh is making fun of me here is that, you know, I've started building my own just internal tools versus leveraging external tools for like financial management or, you know, task tracking, things along those lines. I just have my own stuff running and it's probably not as fancy and have all the bells and whistles, but like it's very fine-tuned for what I need in life versus shoving into somebody else's system. And who knows what everyone's doing with these training data. So at least, you know, the, the big companies, you know that they're already stealing it. What Anthropic is doing with all the data you put in it, I assume they're stealing with it, stealing it at this point. But like, I know Meta is, like, I know that's what Meta does. Anthropic, they're probably doing it at some level.

28:27 Yeah, I mean, you'd like to think that Anthropic is, you know, consuming the metadata specifically to, you know, understand their own usage. And I mean, there's, there's built into the contractual language of the Pro and Max subscriptions about the, the training on your data.

28:42 Mm-hmm.

28:43 But, you know, there's a lot of data that you're sending alongside, you know, with that data that you don't, you're not even aware of. But it's more, you know, like, I don't know, I'm more worried about like when I have agents running inside my house, like I'm, I'm in control of the egress to the internet. And running it in, you know, something else, I just don't have the same visibility and control. It's probably much ado about nothing, but it's just something to be considered of. Like, I'm not really worried about the prompt including my password necessarily, 'cause I don't think that's really gonna go anywhere. It could, you know, it's gonna live on some Anthropic system and that's not good, so I wanna avoid it, but it's not what keeps me up at night.

29:22 Well, they tell you if you accidentally paste a password in. I did something that looked like a AWS key and it wasn't. And it's like, this looks like a secret key. You should re-change this. I was like, thank you, but you're wrong. But I appreciate you telling me and making me double guess, double check that I did not do something stupid, like give you my access key secret key.

29:40 Mm-hmm. All right. Anthropic is just donated its agent skill standard to the Linux Foundation, Agentic AI Foundation. And that matters for anyone building or running agents across multiple platforms and clouds. Agent skills is an open standard where a skill folder containing a skill.md file, plus any scripts and reference material, teaching an agent a specific text or domain. It originated in Claude code to supply the procedural business knowledge and general purpose knowledge that agents lacked. Adoption is super broad, over 70 different agent platforms, including ChatGPT, Kurscher, GitHub Copilot, and VS Code have integrated it into their systems since December 2025. And over roughly 1 million skills have been published and nearly 280 million installed. Anthropic is handing over the name, logo, GitHub organization, and agentskills.io domain to the AIIF, which it co-founded with Block and OpenAI. This follows an earlier MCP donation, so both major agent interoperability standards now sit under the same neutral foundation. Anthropic, Google, OpenAI, and others will maintain the standards as individuals with David Soraya Pera, and Mahesh Murug as lead maintainers. The proposed roadmap includes version releases, a formal proposal process, and security guidance for skill authors and installers.

31:02 It's great to see that these are getting kind of more uniform across all the platforms. You know, there was Agents and Claude.md and, you know, like, like the, the ecosystem even among the vendors is starting to kind of standardized a little bit, which is great for, you know, somebody that's developing because, you know, there's times that you definitely don't run out of credits on one tool and go to the other tool and having everything kind of link up so nicely together is good. And having these standards that they're talking about out there from security, you know, will help everything become slowly be a better and just more secure world. You know, as we talked about early in the show, Security is just so prolific now across everything that like having that be standardized will just be good for everyone.

31:51 Yeah, it'll play nice like into the harness ecosystem. Everybody's really pushing, you know, you can really make these models more swappable, you know, as they change. And you know, like the security aspect, it's now, you know, like you can get your Terraform modules that Microsoft maintains so you have some trust in it. Same thing with skills, right? Like this is the, you know, I'm not going to use it, but you know, let's say there was a payment processor.

32:18 Yeah. I mean, it's, you're absolutely right. Like that's, it's actually a big part of my day job right now is figuring out the strategy on how we're going to secure things like this, right? It's easy to install these from public sources. It's easy to copy and paste an entire GitHub repo and just point your agents at it. I've seen people do it. I definitely experimented like that a few times and you want some sort of process to validate that there's no nefarious instructions in there. And do you really, I don't really expect, you know, the users I'm trying to protect to go through and read everything line by line of, of in those repo. If you're talking about, you know, hundreds of skills and tools and everything, like it's a lot. And so trying to figure out how to provide those safety guardrails is important. So one of the ways you can do that is by trusting this donation, right? Or this foundation rather. And you know, 'cause you know, their, their reputation depends on it and it's built into sort of the quote unquote product that they're offering. So it's very cool. And moving back into the world of AI models are just hacking everything. Anthropic reveals that Mythos-5 tried submitting malicious code to open source projects. Anthropic and the UK AI Security Institute disclosed that Mythos-5 tried to push malware to a real GitHub project during a cyber evaluation, which is a useful case study on how frontier models route along their own guardrails. This was during a simulated corporate network breach test. Mythos-5 followed its rules and initial— its initial attacks failed and then pivoted to a supply chain approach where it compromised a GitHub repo where one of the target servers periodically pulled from. The model created a GitHub account, submitted a malicious pull request, and when the maintainer rejected it, it rewrote the malware and used a second account to pressure the maintainer. The PR was never accepted, so no users were actually impacted, and the run ended when the token budget ran out.

34:08 So, I, I mean, the, the gist of this whole thing is what Ryan said, like, this only failed not because the actual attack vector failed, but because it ran out of tokens, ran out of money. Yeah, like legitimately, like it created two malicious users and then tried to convince the human that they are wrong on the pull request to accept the malicious code. Like it's a crazy complicated attack vector that it did.

34:34 I mean, it's social engineering, AI-led social engineering.

34:38 It's ridiculous.

34:40 Well, bots have been trolling forever, so I don't know, this just seems like the natural, like where things are going. I don't know, this attack is not as impressive as some of the others. Like, yes, it's upsetting, but also it just seemed like the natural progression for things there.

34:54 Well, doesn't GitHub also have like, are you a robot on GitHub setup? Like, that means you need email addresses and things along those lines too. Like, it had to make its way through GitHub once, fail, think about a separate way in, redo another PR, create another user, At least when humans did it with like, what was it, the XV one that, that was found because, uh, the Microsoft saw like the lag of a couple seconds, like milliseconds in PostgreSQL connection.

35:26 Oh, the LZ.

35:27 LZ, is that what they're calling

35:28 it? Yeah, yeah, the Linux library.

35:31 The attacker had been in the, in the environment, you know, and been, you know, got to the point of being a trusted, excuse me, trusted provider for like, you know, it took like a couple years to get there. Here it's like in a matter of hours or days, I don't know what the timeframe was for this, didn't go into that many details, but like it created a second user and was like, no, no, human, you're wrong. You should accept this malicious code.

35:56 Yeah. No, I mean, it's, I'm, it is just sort of crazy. You know, it, I'm hoping that the accounts looked false, like there was no, you know, other commits or something. 'Cause like how deep did the AI go? But, um, it's only a matter of time before, you know, more sophisticated attacks like this are observed, right? Because it's, it's not a new method of attack. You know, humans have been doing this exact same thing forever. So,

36:21 I mean, I'm more curious, we're hearing about these types of attacks, what aren't we hearing about?

36:30 I mean, there's— I know there's something, but I also feel like, you know, what we're hearing about are the models where they're specifically not putting any safety guardrails. So these aren't You know, this isn't a model that, you know, a normal person can use, right? They can't query this. It's built into their, is it Glasswing or Icewing or whatever? They're, they're a separate, you know, uh, thing. And so it's, it's part of those security research, you know, engagements, which is cool because, you know, as a security engineer, it's, I've definitely ran into several times where I ask a security question. It's like, oh, I can't do that, Hal or Dave. And it's just, you know, like, 'cause it's the security tooling, but you know, it is kind of interesting in the sense that, you know, this is exactly why we're building these tools is to find, suss out vulnerabilities and, and do these attack chains. But it also, it's funny that we don't like it when it happens. Like, oh no. Oh, it's 'cause it's dangerous, right? Like, it's scary.

37:26 Yeah. It's scary what we've gotten computers to do. I think is kind of the principle of it. Like, we've gotten them to watch how humans attack and then do that automated and quicker and better than most people can do it.

37:44 It's their reasoning and like the hyping themselves up, like, now this is right, I'm in the right, like, like they're building that, you know, when you look at some of these attacks and they're like, oh yeah, the, the chat had some sense of like morality and was convincing itself of it was doing the the utilitarian thing. I mean, it's all math under the hood, but it's creepy math.

38:06 But, you know, it also could be a marketing stunt for the upcoming story.

38:09 Well, yeah, nice segue. So Anthropic files the S-1 with the SEC for proposed initial public offering. So it's— Anthropic has filed for its IPO, a notable milestone for one of the major frontier AI model providers that cloud customers depend on. Anthropic filed its S-1 for Class A common stock planned for the NASDAQ Global Select Market under the symbol ANT, or Ant, uh, which I'm surprised is not taken. Share count and price range have not been set, and the registration statement is not yet effective, so no securities can be sold or offers accepted yet. Goldman Sachs and Morgan Stanley are lined up and saying shut up and take my money. I mean, they're the sources for the preliminary prospectus indicating they are leading the offering. The public S-1 will include audited financials, revenue figures, compute commitments, and cloud provider relationships, which the hosts can dig into once the full filing is available. And yes, we will. Um, that is definitely something I'm interested in.

39:09 So I've gone through it. There's another article I have, um, that I looked up and, you know, it called out a couple interesting things in the S-1. One is around their culture and the retention. And I'll post a link to the one that I'm referencing in our show notes too. But they are retaining customers. And even among the, every customer I call, like, you know, or anyone I talk to during my day job, it's, they're on Anthropic, you know, by far is the most. And then followed by, you know, OpenAI is probably the second, but like, it's a fair way down where they are. The second kind of level that interesting point from here was the enterprise adoption. You know, I feel like Anthropic took it by storm with the enterprise adoption, and I don't know if you guys kind of think the same, but I feel like most people and a lot of businesses moved to Anthropic before ChatGPT really got everything moving forward on it. You know, so like that was kind of one of the other big ones that I thought was interesting. The other part of the, that I called out here was inference costs has come down over the years. So while they're spending more, they're doing more with it. So like they're not seeing, yes, it's going up, but like they're getting a lot more revenue for every dollar they're putting into training too. So the percentages are a little bit different.

40:33 Yeah. I mean, that makes sense, right? Because you're seeing, you know, all the token caching and all that stuff. And that's all of that is going to bring inference costs down, which is why, you know, your subscriptions go further, which is good. Yeah. I mean, it is interesting 'cause isn't Anthropic, they're not quite a not-for-profit like OpenAI is, but they're, they have, what is their, they have a different sort of class of something. Oh, public benefit corporation. And trying to figure out what the impacts of that are for for, uh, being publicly offered and traded on NASDAQ is interesting. Like, I don't—

41:12 I feel like we need Justin for that. He'll know all those nuances, differences.

41:16 That's the business side. Yeah, for sure.

41:18 Where's our executive when we need one in the, in the room? You know, one of the other things I saw people online talking about was, um, it was called out as a risk in the S-1 is the supply chain of like hardware. Is a concern for Anthropic and like, oh,

41:34 for sure,

41:35 it makes complete sense. But like, it is, like, it is called out in the S-1 that that is a concern.

41:42 Yeah. I mean, if you can't run AI queries against it, then you know, you're not gonna get subscriptions either. You know, there's no revenue.

41:52 Yeah. I didn't dig into, there was someone that posted, I ran outta time before the show as I was reading up with some of this was around, they were saying that there was some interesting stuff about where their committed spend was, wasn't quite where everyone thought it was, but I honestly didn't have a chance to finish reading that.

42:06 So, oh, I mean, this is something that we're all very interested in on the show, right? So it's maybe Josh is it, 'cause you know, he's looking a little glassed over on the thing awaiting. But just kidding, you know, I'm sure we'll be talking about this more 'cause it is like, you know, the first real big test, I think, you know, 'cause Gemini is a Google product. It doesn't subject to the same risk. And OpenAI has delayed their filing and is in a kind of a weirder state than Anthropic is.

42:34 Yeah. I thought they pushed theirs to like 2027 'cause they did their S-1 early and then they hit the pause button and they pulled it.

42:39 Yeah.

42:40 So I don't really know how all that works though. Like I think like it's like 4 to 6 weeks they're at, or 4 to 8 weeks like from now is when they expect to be public. So Hmm.

42:52 I don't know if that's set, 'cause it did seem— I've read a couple articles where it's like, and the— where's the date? We don't have a date.

42:57 We don't know which.

42:58 Yeah.

42:59 Yeah.

42:59 That makes more sense.

43:00 It's, you know, this has definitely been an interesting year for IPOs. You know, SpaceX, you know, got Anthropic now. It's something to tell the grandkids about. Where were you when this one?

43:12 That's true. Yeah. I'm not counting Groq. With the SpaceX, I'm not counting that.

43:18 Uh, I, yeah, that whole, that's a, that's a tangent right there. Those business entities and how they're intertwined. And, but, uh, yeah, I think Anthropic might even be involved in the compute that they're trying to spin up over there, you know.

43:30 That's what the article was that I have time to finish, is like they have committed, but like they're only committed for like a couple, like months for each one of these things in a lot of places. And so it's, you know, somebody did a research you know, and has an article. I'll try to find it. I honestly, I can't, I wasn't, couldn't, I didn't finish it 'cause I couldn't find the article again. But it was like kind of somebody trying to overlay like news stories to what was in the S-1 of where committed spend was versus, you know, what they publicly talked about and seeing kind of what actually made sense.

44:02 Okay, and our last AI story, which I just touch on briefly, which is that OpenAI has announced GPT-4. 6.1 tera. It's a faster, cheaper model aimed at computer use and visual reasoning, and it's worth a look for anyone running high-volume automation. It's available in all the places you'd expect an OpenAI model to exist, and it scores 15 points above GPT-5.6 tera at about one-third the cost and matches the GP6 Astra at what, one-tenth per token price. It's proved notably with the error rate, um, on error-inducing prompts, jump dropping 53.6%, which is a lot. And the coding results nearly doubled with Terminal Bench going from 21.5% to 50% at max effort. Again, practical use case for just the high-volume back-office automation such as reading invoices, entering data into accounting software where the cost per task drop makes agentic computer use more economical.

44:57 I mean, it's great for them to release this 'cause this is equivalent to like Haiku, right? If I remember correctly.

45:03 I think so. Yeah. That's what, that's the way I read it anyway.

45:06 Yeah. That's what I understood. So like having that a new model, you know, like Haiku 5.5, I, I find exponentially better than 4.5, you know, which makes sense. It's one year later, you know, and I kind of see the equivalence here with this where it, it's going to be making these things better cuz the backend office admin work is where I think a lot of AI can really make it a big improvement. Improvements.

45:29 It's definitely the area I need to flesh out in my own workflows, right? Like I'm using it for more reasoning tasks and thinking tasks and, and bigger projects, but I really do need to look at like the admin things that I do and really replace it.

45:41 Expense reports.

45:42 Yeah.

45:43 Fantastic for expense reports. You literally just, like, I just give it to Claude. I'm like, here's all my expense reports, go put it into Concur, I think we use, and just do it for me. Though it did duplicate one, so I have to deal with that after the show.

45:58 My problem, like everything AI, is Signal. So it, that requires me to be organized enough to have all my stuff in a place where I could point AI at it, and I'm just not.

46:08 Have AI do that for you.

46:09 Yeah, that's what I need. I need AI to do everything.

46:13 You need a project to prep you for the next project.

46:16 Mm-hmm. Well, onto other cloud tools. Introducing 1Password for Agents. 1Password announced 1Password for Agents with a brilliant naming convention, a way for personal AI agents to log into sites on your behalf without actually seeing your credentials. The problem was before, when an agent hit a login page, it would stall, force you to take over the session, manually copy and paste or type your password in if you're a crazy person and still knows how to type passwords, and expose them to the agent, which exposed them to the agent. This current approach has the agent request credential from your Vault. You approve the 1Password, 1Password automatically fills in the destination site so that the password or the one-time code never passes through the agent. Launch integrations with Muse, Claude Duet, ChatGPT, Work, and Dots. Grokbot, TownTownie, and multiple others are coming soon. Rollout timing and the platform is varying by each individual partner. I am very happy for this. Just saying.

47:18 Yeah. This is a, like the, you know, talking about Duet and adopting that kind of workflow. One of the things that's prevented me from doing that is that for some reason, whenever I try to use one of these tools and it requires me to still be involved, like the amount of rage I feel is not proportional. And so, and passwords are a big part of that. Like, go pay all my bills and then you know, the OAuth session has expired, and it's like, I got to go log into these 17 different sites to pay all the bills, which I would have had to do if I was doing it by hand, but now I'm mad about it. So it's really funny.

47:54 Yeah, just feel like that, get out of my way, I can do it on my own.

47:58 Uh, my problem was I didn't realize this was released. I hadn't seen the news drop, and all of a sudden I was doing something and 1Password popped up, and I was like, Wait, is this a— but like, did this get hacked? Like, it might have like some malware on my computer, you know. I'm unsure kind of where this ended up, so I got a little worried at first. You know, it's the fun of, you know, things auto-update, which is great, but they get new features, and the security person to me is like, hold on, what just happened? Versus, you know, knowing that no, no, it's fine, it's from 1Password. So it is a nice thing. I did play with it a little bit. It does worked nicely, at least with, uh, Claude, from my experience. Claude, uh, just normal Claude. So this is technically called CoWork, but whatever they want to rename that today.

48:43 Oh, they'll change it and rebrand it and combine it with something else. All right, moving on to AWS news. Amazon's cloud capacity is sold out through 2027. What? Andy Jassy says the capacity is largely sold out through 2027, with commitments reaching into 2028. Which is kind of crazy if you're planning an AI workload on AWS. According to Jesse, the 2026 capacity was already sold out long ago, and a good chunk is already being sold and leaves some portion of unreserved for on-demand EC2 and Bedrock inference, but the large blocks are spoken for. For customers, this suggests securing your GPO and accelerated capacity for large AI training and inference projects is super important and may require multi-year commitments made well in advance. So, you know, prepare your wallets. Amazon has guided to roughly $220 billion in capital expenditure for 2026, and Jesse said the company could invest substantially more without it being a poor use of capital supply, not demand, but it appears to be this constraint. Demand is described as coming from both AI labs and enterprises, which points to the capacity competition between frontier model builders and typical enterprise adopters worth discussing how this affects pricing.

50:01 I also feel like this is a way for them to sell more PPAs. Like, we don't have capacity, but if you sign with us, we'll guarantee you this capacity, which like, I guess isn't bad, but I also know that they can't, you know, no one has capacity right now out there. So, you know, people are throwing money saying, I need this capacity. And I'm wondering how much of it is just sitting idle or not used, which I hope not a lot, you know, because people have thought they were going to use it and then didn't end up needing it or kind of how Amazon's managing that on the backend. You know, Ryan here said I need, you know, 5 GPUs and Matt said he needed 3 and I'm only using 1. So Josh comes in and requests 2. Are they actually available to the random on-demand user or not? I don't, I kind of wonder how they're reserving it on the backend.

50:51 Yeah. I mean, it's an odd thing and it's like they, if stuff's already being sold into 2028, if you come in and want reserve capacity, do you, do you have to wait until 2028 like to get that reserve? Like, will they even sell it to you? So that's, it is sort of crazy. And it's Amazon, like Amazon's not, not never the cloud that has the capacity constraints. Like I'm used to this from Azure and GCP where it's just like, this isn't very cloudy, but when Amazon comes out and announces that the AI hardware constraint is real.

51:24 Whatever happened to the Chinese model that was able to train their model on more generalized compute and older GPUs than requiring the newest ones? Is that still—

51:36 it was a DeepSeq training exercise, I believe. Like it wasn't anything special, but it's like, it's, yeah, it's achieving the similar level benchmarks with building that model on way different hardware. But according to OpenAI though, that's a lot of distillation of the models, of the GPT models and the whole thing. So like there's a lot of like nuance in there that I don't know which is real, which isn't.

52:02 Jonathan, come save us.

52:04 Yeah, exactly. All right. AWS Network Firewall now supports Route53 resolved IP validation. The firewall can now validate that your connection's destination IP was actually resolved through your VPC's Route53 resolver, which matters if you want real domain filtering beyond just web traffic. Resolved IP validation extends domain-based filtering past HTTP and HTTPS protocols, and so you can use it on things like SFTP and SSH, and teams can write down domain rules for any outbound protocol instead of relying on brittle IP-based rules. It closes a bypass gap by checking that the destination IP came from the VPC's route resolver, and clients can skip DNS or hardcode IPs, can't sidestep domain rules. It pairs with Route53 Resolver DNS Firewall for layered DNS and egress security, and a useful discussion point is important for anyone who wants to do defense-in-depth outbound filtering. This is available at no additional cost and is available in all commercial regions where Network Firewall and Route53 resolvers are offered.

53:14 Ryan, you stopped my way to break around your rules. Using tools the wrong way is what I like to do at times. And this sufficiently actually validates stuff, which is good. You know, it's not an attack vector I would normally think about, but it is something that they, that is a good fix, you know, and it does, you know, it's a belt and suspenders model essentially. This is the suspenders now set up. So.

53:40 Yeah, it is kind of interesting to think about, like, you know, egress filtering at layer 7 is pretty common where you have, you know, either you're using categorization or specific domain listing, but then if you could figure out what those rules are and you sort of use Etsy hosts or do something, you know, along those lines to sort of bypass the DNS of it and get past that filter. You know, it's kind of interesting. And then, you know, for SSH and SFTP, that egress filter, you know, you can inspect the layer 7. So this is kind of an alternative to that, which is neat.

54:11 Yeah. Nobody would ever just put like 8.8.8.8 in their DNS on, you know, a corporate system for any reason.

54:17 No, no. Why would we do that?

54:20 We've never had to troubleshoot that way.

54:22 Definitely never had to debug something along those lines.

54:25 Uh-huh.

54:25 It is always interesting to see like the best laid security plans fall to like some decimal entered somewhere, you know?

54:33 Mm-hmm. All right, moving on. AWS announces the AWS Interconnect Last Mile preview with Megaport. AWS and Megaport have teamed up on a preview of AWS Interconnect Last Mile, which aims to make connecting branch offices and data centers to AWS a few clicks in the console instead of a multi-week provisioning exercise. Serviceless customers pick an AWS region, the amount of bandwidth, and the Direct Connect gateway, and the partner's Trevor ID directly in the console, and then you can use an AWS-generated activation key to finish provisioning directly with Megaport. AWS pre-provisions capacity and automates BGP peering, VLAN configuration, and ASN assignment, removing much of the manual network setup that makes traditional Direct Connect slow to deploy. Megaport brings SDN spanning 1,200+ enabled data centers in 31 countries available at launch, and customers will be able to scale bandwidth from 1 gigabit per second to 100 gigabits per second dynamically directly in the console. The service is backed by a 99.99% SLA with zero downtime maintenance. And while the preview is only limited to San Jose, US West 1, and and US Tire Fire, US-East-1 metros can then be created in those areas via console or CLI or API.

55:56 I never realized US-West-1 was actually labeled as San Jose.

55:59 That's the first time I've ever seen that. Yeah, I thought it was SFO.

56:03 Yeah, I mean, I know they're in the, like the code, like if you got like the logs from Amazon is that, yeah, it's listed as Northern California.

56:11 It's probably more than just San Jose considering the size of the, the Bay Area.

56:16 Yeah,

56:16 but it's probably a data center in Fremont, a data center in San Jose, a data center, and maybe San Francisco Direct. Who knows?

56:24 I mean, has anybody ever had to set up an interconnect before? Ryan, I think you have had to do it before. It's a pain in the butt, to put it nicely. Having this, you know, if you need a, especially like a smaller segment setup, or you already are on Megaport and you just want to link up your AWS, is fantastic. It just makes life a lot easier. So, you know, it's a nice quality of life improvement if you are having to do this decently often. But I've had to do, I think, one or two Direct Connects, and I've done one or two more than I've ever wanted to do.

56:57 Yeah, it's a very non-cloudy experience, or at least the previous experience was very non-cloudy, right? It's like you want to get into a console and provision a network and have the connectivity set up and, and get to work. And then this very quickly devolves into the older data center model where it's like sending an email, tracking a ticket through a system, communicating that, going back and forth with capacity constraints and design things, and it sucks. And so I'm really happy to see this be fully automated in the console.

57:23 Did you ever have to deal with the original HSM, like HSM v1 on AWS, where it was like a two-week, like you had to open a ticket to request it and then it took a day. And you have like 2 weeks to use it and then, and it's ridiculous to us.

57:39 I think, uh, I was warned off of that 'cause I would've never done that.

57:43 Yeah, it was bad.

57:44 We, and I definitely didn't have to. Yeah.

57:46 We did a project for a customer once where we like, we're like, okay, we're spinning it up. And if you like cross that like 2-week mark, it was like, you know, $5,000 or something.

57:55 Yeah. I'd buy one and run it under my desk before I would do that.

57:58 Yeah.

58:00 All right, AWS Transit Gateway and AWS Cloud WAN now support a one-to-one NAT. Hooray! This finally gives teams a clean answer to overlapping CIDRs without renumbering or running multiple NAT appliances within your VPC. This defines address translation rules directly on Transit Gateway or Cloud WAN covering VPC site-to-site VPN, Direct Connects, SD-WAN, all through Transit Gateway Connect. Translation applies automatically in both directions, so there's no return traffic rules to write and no NAT appliance to deploy, patch, scale, manage, love, cherish. And the main use cases are mergers and acquisitions, partner network connectivity over VPN, multi-cloud links, or interdependently provisioned VPCs that also happen to share ranges like 10.0.0.0/16. Never once had that problem, never. Setup is available through console, SDK, and the usual methods, and pricing was not part of the announcement.

58:59 I assume it's just normal networking costs. They're not gonna charge anything extra for this, but I could be wrong. This feels like a decent, you know, quality of life improvement. I haven't had to deal with this luckily. I know I almost did once, but I've talked to people very carefully when they set up their initial networks and I'm like, okay. You have a VPC. While in theory everything in the cloud gets created and destroyed, rebuilding your VPC sucks. So let's make sure you set it up correctly and had a lot of upfront conversations with people about IP schemas. Because it's one of those things that everyone's like, it's not a big deal until it is. And this is one of those situations that it sucks.

59:41 Yeah, when you get that like merger acquisition though, there's nothing you can do. Like if you both sides did everything right, then you'd probably have a collision still.

59:48 Mm-hmm.

59:49 Just, uh,

59:51 and your only option was to route everything through something that could do that translation for you.

59:56 It's, you know, old problems are new again.

59:58 Mm-hmm.

59:59 Like I dealt with this on-prem, but it's the exact same experience.

60:04 Yeah. We've just moved our data centers into the sky.

60:07 Yeah. I had something to hit with a bat afterwards, you know.

60:13 AWS has announced the general availability of AWS Continuum for code vulnerabilities. It's an AI-driven service that finds, validates, and patches code vulnerabilities, which matters to any team that is drowning in unprioritized security findings, which is everybody. Continuum builds a context graph from your AWS environment, resource topology, identity relationships, network reachability, for example. And so it can tell which findings carry real business risk rather than just ranking them by the CVSS score. It ingests existing vulnerability findings and scans code, then tests the findings in an isolated sandbox to confirm the exploit. This aims to cut false positives and focus engineers on issues that can actually be exploited. Remediation arrives as pull requests in developer workflows with the exploitability proof and blast radius assessment attached. It selects the AI model best suited for each task, and adoption is staged. Learn mode will propose the human approval and force mode will act completely autonomously. And you can set up guardrails you define once you trust its precision. It works alongside AWS GuardDuty and AWS Security Hub rather than replacing them. And while the announcement does not include pricing details, it's going to be crazy expensive.

61:30 Yeah, it'll be interesting. It'll, it'll be token-based or what. But, uh, I do love that these new security products are not just like chucking CVEs over. They're, they're, hey, here's some remediation path, which really helps ease that, you know, as you're reaching out to engineering and development to like, from a security perspective, to get things done, uh, you get a lot less pushback.

61:53 So they have done pricing. It's by lines of code, $200 per million. For the first $10 million, then the next $90 million are $180, and then it tiers down from there. I'm not gonna read everything. It's on, it's on the pricing page. I did look.

62:11 That's cool. Yeah, it just wasn't part of the answer when we wrote the show notes. That's cool.

62:14 Yeah.

62:15 I wonder how that compares. Remember that service CodeGuru, like, which was just cost prohibitive to run?

62:20 Did they deprecate that?

62:21 I believe it's gone now. Yeah. 'Cause no one used it 'cause it was too expensive.

62:25 It was ridiculously expensive. I ran like small things through it just outta curiosity to play with it, but It was expensive. Is it EOL?

62:34 Yeah.

62:35 Yes. CodeGuru Security is end of life. Announced retirement on November 20th, 2025.

62:45 But I mean, this is great. This has been, you know, a decades-old problem where, you know, security teams generate the list of vulnerabilities scanned in an environment. They toss this giant Excel spreadsheet full of hundreds of thousands of things to address and and dev teams and infrastructure teams and cloud teams just look at it and be like, what?

63:03 How many of us have all sat on a call with somebody pulled an Excel spreadsheet open and said, okay, we're gonna start at line 1. And you're like, there's 27 people on this call. How is this beneficial to everyone? Like, so any level of prioritization, you know, is important. And I know prior job, I always tried to go and actually give people real priorities versus, there's a low in something that doesn't have an update. Cool. What do you want me to do with that? That it's a low in something that isn't fixed and it's Microsoft's fault. Like there's nothing I can do. Microsoft's not updating it. Great. Let's accept and move on and exclude it for 3 months until we actually can have a conversation about how we're gonna re-architect the entire platform.

63:47 Yeah.

63:47 You know, but that's a different—

63:49 chasing people to patch things is not effective, right? So in my day job, I focus for vulnerability management. I focus on, you know, building platforms and services that make patching either automatic or, you know, easy button or, and just focus on the, the components you're launching into that infrastructure, have them already be patched and having some sort of mechanism under the covers. And, you know, I, I like that this, if you layered something like this on top of it now for everything that's running and, you know, everything that's sort of, there's a vulnerability that's discovered over time. This could find it and then offer at least a PR to address it. Doesn't mean someone would actually build that into the codebase and deploy it, but now you can at least chase them down for that.

64:29 How is this gonna play with Inspector? Is it just, uh, Inspector runs on the running instances and then kind of Continuum monitors, you know, AWS and Inspector? Or like, I don't, I haven't thought about enough about how those two interconnect at this point.

64:45 Well, I mean, there may be interoperability under the covers, we don't really know, but I mean, if it's, I think it's largely looking at deployment code and infrastructure code to determine the vulnerabilities from there, but then using the sort of digital twin context graph of your AWS environment to determine the risk score. That's how I read it.

65:05 That's what I just found when I was researching. It doesn't replace Inspector. Instead, it consumes the finding from Inspector, GuardDuty, Security Hub, and adds AI reasoning and deep penetration tests along with a bunch of other stuff on top.

65:16 Yeah.

65:17 And produces the pull request. So it's a add-on.

65:20 Yeah. I mean, this is the right model. You have to do context-based vulnerability remediation with the amount of vulnerabilities these days, especially the amount of vulnerabilities that are being introduced, but because AI is finding them all. So. All right. Amazon. And you can now run EMR on EKS and it reduces your EMI charge by 30%. That's pretty much all we wanted to say about it. There's more details in our show notes, but we were, most of it was me and Matt going, oh, you could run MapReduce on EKS? Huh, I never thought about that.

65:55 I mean, it makes sense. I just never knew that that was a service that they had to run on EKS. Yeah. Nice, it's cheaper.

66:03 Yeah, 30%. Can't argue with that.

66:06 Just don't tell the CFO, use, use the money elsewhere.

66:11 Oh, he's just going to get the bill anyway. Like, yeah, don't worry, I'll spend that money elsewhere. AWS Direct announces a 100-gig expansion in Bangalore, and the only reason we're really including this is because that's where Justin is, and that's why he's not on the podcast. And so he promises he has nothing to do with this, but, uh, I don't believe him. And, uh, so I think we can all thank Justin. He's added the Direct Connect capacity by hand, presumably shuttling the packets personally.

66:38 Um, he's physically wiring it for everyone, just so you know, making sure that the tubes are cleared and it's good to go. He got it up to 99 gigabits remotely, but he wanted that last 1 gigabit in person,

66:49 you know, and he's thorough, you know, and so now, now he's there. And so. Hopefully he'll be done shuttling packets soon and he'll come back to the podcast so that me and Matt have less work to do. I mean, so that we can deliver a superior show.

67:05 I feel like Josh should be arguing right now a little bit about that, but that's fine.

67:10 No, no, I've listened to the show. This is spot on.

67:15 So how's the show differ when it's Ryan and I versus Justin's here to, uh, guard all of us and You keep us on track.

67:23 Yeah, well, there's definitely more tangents.

67:26 You ever remember being like in middle school and there was a substitute teacher?

67:30 Yeah.

67:30 Are

67:31 you the substitute teacher? Today,

67:35 yes.

67:35 I have snacks, I will be playing a movie. Please don't set anything on fire.

67:41 Yeah. Don't die, that's my only request.

67:47 AWS ECS Auto Scaling now supports launch-before-terminate at the account level. So Launch Before Terminate is now an account-wide default for ECS Auto Scaling, which means fewer capacity dips whenever instances get replaced. Launch Before Terminate can now be enabled once. Any Auto Scaling that launches a replacement will wait for it to pass health checks before terminating the old instance instead of configuring each Auto Scaling group individually. It covers instance refreshes, rebalancing, health check replacements, and maximum instance lifetime replacements, which are the main events where groups previously risk a temporary drop below the desired capacity. This is available in all regions at no additional cost. And I have never seen something like this set at the account level before in AWS, to my knowledge.

68:34 No, I assume this is there. I don't think most people ever are going to care about this if you're doing auto scaling correctly, but I've seen definitely a couple of times where people are like, they have two and they replace one and then the application goes down and then the first one goes down. 'cause they don't have auto scaling set up to really what they need is 3. And I assume this is just set up because Amazon's sick of support tickets about stuff went down and they don't want to deal with it. So here's an API for your center of, for your COE, your center of excellence to actually go solve this at a global level for all your teams that don't actually know what they're doing, but it will make it actually work.

69:10 I'm just curious the math of like support ticket cost versus like the compute costs. They're just like, it's a no-brainer.

69:17 Well, the customer pays the compute cost. It's more implementation to us.

69:20 Yeah. This just makes sense, right? It's, it's a couple minutes of extra runtime. This is, that's probably the real reason why they're introducing this.

69:27 Yeah.

69:27 No, probably not.

69:28 It's like the toothpaste tube, just get 30% bigger, you know?

69:33 Mm-hmm. Did you add more toothpaste in there?

69:37 No. Yeah. AWS Identity's IAM Identity Center now supports cross-region trusted identity propagation. This allows users' identity to travel with their requests across regions. Previously required all services using a user identity to be in the same region and you had to reauthenticate when moving across region, which is awful. Can speak to that firsthand. The previous workarounds for broad IAM roles that couldn't distinguish between individual users or consolidating everything into one region, which is obviously bad.

70:11 Well, then you had issues with DR too and failover. And if your primary region where Identity Center failed over, you, or went down, 'cause I don't know, you launched in US-FIRE-1, then, you know, when you tried to failover to US-BOST-2, you had other problems. So, you know, you couldn't log in, minor details like that. So this is a great quality of life improvement, especially for enterprises that are doing any sort of DR testing You know, it does fix that real-world gap of things that went down the last time US East 1 went down for people.

70:44 Amazon ECR now lets you backfill replication for images that already exist in your registry, closing a longstanding gap that anyone running multi-region container workloads has probably hit. Until now, ECR replication only applied to images pushed after the replication rule was configured, and so teams had to manually retag all the existing images and force them across the new repo.

71:08 So I ran into this again in a DR situation where we were, we're in a real DR, but where we were testing DR and we had set everything up and this was post our prior one. And we had this one container that we just hadn't ever pushed an update to, which clearly means we weren't patching very well at the time. And it didn't get an update and then it couldn't boot up and it, hindsight vision, it wasn't there because it wasn't set up in advance. So this is a nice, you know, you've done a thing, you, you know, you built out your initial environment, you've then run the, you know, you've got it up. Now you're going for your SOC compliance for the first time and you set up your replication rule and you realize that you haven't actually, don't have them actually in the DR region. So it's a great way to just backfill it. Again, great way for my, for Amazon to make money too.

71:57 Kids don't know how good they got it, man.

72:00 Well, that's why you test your DR process. So kudos to you, man. Yeah. Amazon ECS now supports GPU sharing with NVIDIA Multiprocess Service. So now you can let multiple ECS tasks run on a single GPU, which matters for anyone paying for underutilized GPU instances, which is everybody. and especially since there are no more GPU instances until 2028. ECS can now place multiple tasks with the concurrency on one GPU using the multi-process service, and you declare the memory allocated for each container and cap its total share of compute. None of this requires code changes, and the target workloads are those that only use part of a GPU, obviously. So small model inference, embedding generation, multimodal serving, that sort of use case. On ECS managed instances, ECS monitors the MPS control daemon on the instance, and if it stops or becomes unhealthy, ECS will automatically replace that instance. Setup involves a capacity provider with NVIDIA GPU instances and GPU sharing strategy configured in the task definition. This is available in all AWS commercial and govcloud regions, and no separate pricing was listed throughout the announcement. Because I imagine it's just quality of life improvement for anyone who's using containers to train models or run other inference.

73:21 I mean, I like to gain more out of my current things without paying more, so I'll take it. I'm more curious what they changed under the hood to make it work, but you know, it's more of their problem than mine.

73:33 Probably pretty much the same as like what Databricks did, right? Everybody's relying on these new features and they're like plugging it in. Which, which is great.

73:41 But I thought Lambda runs on their own microVM, the Fire— what's the Fire one? Fire— not Firebase, Fireworks, Fire—

73:51 Firecracker.

73:52 Firecracker. Thank you. I was getting there. Framework. So I guess maybe they then run it on top of Kubernetes. I don't know.

73:59 Yeah. I mean, even if it's their own stuff, bare metal and orchestrating, like they're gonna port that technology over.

74:07 Yeah.

74:07 I probably just had said AI, go do this Claude.

74:12 No, they said Kiro, go do this.

74:14 Oh, right, right. Kiro.

74:17 Onto the world of GCP. Google Cloud brings OpenAI models and expanded cloud availability onto the Google Cloud. Google Cloud just added OpenAI GPT-6 families to model garden. 3 OpenAI models are rolling out in private preview, Astra, Sol, and Luna. They've also added 2 new ways to run Anthropic Claude models. Anthropic is a new, is a data processor under the terms and condition on Google Cloud now.

74:47 Yeah, so this is interesting, right? Because they don't let anyone else run Gemini and you have been fairly locked in to a lot of Gemini Enterprise products to only using Gemini models. And now they're expanding that out where you can use OpenAI and, and Anthropic models in those existing sort of harness ecosystems. Like the Vertex AI model garden has other models, um, and this is expanding on those, but the, a lot of the sort of more turnkey platforms, you didn't have any choice. And so if you've got that use case where you need to have all of the frontier models for some reason, there's only one choice for running that now because they sort of control one and they're not releasing it to anyone else, which is kind of funny.

75:36 Wait, so you couldn't run Claude and OpenAI there in the past, or—

75:40 so depending on what tool, right? So if you're building your own Agentic app or an ADK, you could use the models that are available in the, the, the model garden, whatever was available at the time. And so it, it varies like, 'cause it's, they don't, release them at the same time. So if OpenAI releases a new, new model, the Google Model Garden is going to lag way behind on that. And so, but for things like Gemini Enterprise for Business, any of the, the sort of, uh, agent builder platforms where you're sort of running— it's more of a managed service for running AI workloads— you only had the option to, to use Gemini models.

76:18 Oh, interesting. I didn't realize that.

76:20 Mm-hmm. And so this is a big change. And so like, it's, you know, 'cause we're a Google shop at the day job, like this is something that you run into quite a bit. Like, oh, we can't use Claude, you know, for, for this type of use case 'cause we're running it in this Harness ecosystem and there was no other option. And now it's starting to become more of an option.

76:38 Hmm. Well, that's a great quality of life improvement for them to add in. Yeah.

76:43 I also think it's a big, uh, statement on their Argon 4, which is a big model improvement in the Gemini series. And so now that they have sort of a competitive offering, they're like, oh well, now we'll open it up.

76:55 Now that we have something that everyone likes, we'll go for there. Yeah,

76:59 yeah,

77:00 yeah.

77:01 Probably also makes it easier to like bring people that are using those other ecosystems over, right? Like, because the clouds are always trying to pirate accounts off each other. And that's like, there's probably just some hard, like, do you have Claude? Hard no, shut the door.

77:16 Well, from my day job, I've also learned they're changing a lot of the way the funding programs work around, like, on AWS around it, because people were just like, oh, let me go take all your credits, you know, and then just change the API endpoint with some of the programs. So the cloud providers are catching wind of a lot of that and figuring out how to stop people from bouncing around between the clouds just to get the migration credits. So introducing Agent Compute Engine and expanding agent offering. Google is launching Agent Compute Engine, a purpose-built compute service for running AI agents on hardware isolated sandboxes. Agent Compute Engine gives each agent their own micro VMs and separate guest kernel running on the Tanium IPU hardware isolation. It also gains the power of Gemini's agent for work tasks and enter— and it will enter preview soon. You have to sign up for this. The effective design targets an agent loop, micro boot, micro VM boot fast, shared memory via copy on write on a single OS template, and checkpoint and pauses while the agent waits for human or external system. Each micro VM gets a Hyper-Disk for persistent workspace and its own cryptographic identity with short-lived credentials. Nginx injected for egress.

78:35 So this is kind of a neat feature for anyone who wants to, you know, sort of run those agentic operations in like a protected sort of environment, which I, which I like that they're starting to offer more of a managed solution for this, cuz it is incredibly hard to put this together. And so like, you know, it's the, you know, the first thing that people realized when they, when Open Cloud was released, right, was, oh, this has access to everything on this on this computer. And so I, I ran it on my, my laptop and it did a whole bunch of things and now my computer doesn't work right. Or, you know, running any kind of agent stuff where, you know, you have full unfettered access to the web or, you know, whatever system resources you have access to at that local ecosystem. Now you have the ability to sort of run that within these sandbox environments. And there's a lot of tooling that provides a lot of the visibility and security around what that agent can and can't do and allowing it outside. And so this is another sort of thing that's being introduced, just like the, the OpenAI models and the, you know, the offline world. The, all of these protections are going in place because we're seeing those use cases come up where using a model for something, having it go off and do something is having unintended results. And so this is a way that you can sort of hopefully contain it a little better.

79:57 Yeah, I mean, I also assume it's probably some internal software that they've built that they're just kind of exposing to, because everyone's kind of built out these isolated agent work environments. I mean, there's tons of products out there for it too. So I also assume that this is, you know, kind of like AWS does all the time, or I guess Amazon does all the time. They build something internally, they get into their production, they then bring it into the AWS ecosystem.

80:23 I think it says it by different words, but I, it, um, because it's, it says it powers Gemini's agent for work, which is I think what they now call Gemini Enterprise for Business. And so, and so yeah, I think it is.

80:36 Is it Google Workspace or is it Google, you know, quiet you.

80:42 Yeah, no, it's, uh, I hate their naming convention. I always have. So it's definitely something they're, they've created as part of their other platforms. That they're now exposing out. Okay, and finally moving on to Azure, ABFS vector read search for Azure Data Lake Storage. Microsoft just made ABFS vector read generally available, and it's a change that could speed up selective reads for analytic workloads without any query rewrites. The feature merges nearby file ranges into fewer storage requests and runs reads asynchronously, cutting network round trips and per-request overhead for selective reads. It targets compatible analytics engines doing selective reads such as columnar formats, Parquet, and ORC, where queries touch only certain columns or row groups. Engine support is the key qualifier, and so people should check their stack before seeing any expected gains. Obviously, the goal is to have fewer storage requests, and it may reduce transaction costs since ADLS bills per operation, but specific savings will depend on your workload patterns. And there is no separate pricing announced with this feature.

81:50 So this is S3 Vectors for Azure. So something that was announced in July 2025, they're just catching up to? Does that sound about right?

81:59 I think so.

81:59 It's the Azure way.

82:01 Yeah. All right, moving on. Now also generally available is a higher limit for Azure Blob Storage accounts per region, per subscription. Azure just raised the storage account limit per region, per subscription from 250 to 1,000, which is crazy since I, I don't work in Azure, but 250 is pretty easy to hit. So this is a pretty good quality of life improvement for anyone running a multi-tenant or data-heavy environments where you want a lot of different blobs, a lot of different segregation. This is now the default quota per region per subscription and allows customers to consolidate more accounts under a single subscription.

82:40 I mean, for a SaaS provider, this is great, you know, 'cause you can do the way you do encryption if you don't want to do by object, which becomes, you know, more of an application level change is you do it by the storage account. You can't do it by, Josh, correct me, what's it called? The prefix on Azure, the folders essentially. It's not done at that level, it's done at the actual storage account level.

83:02 Container, I think. Can you talk about that?

83:04 Container, thank you. We're recording a little bit later than normal tonight and poor Josh and I are on the East Coast. Still away doing this. So, you know, they don't do it at the container level, which would make life a lot easier. It's all done there. So a lot of companies have had to do a bunch of storage accounts, which also become a pain to manage if you need by-customer encryption key or BYOK or any other feature set like that.

83:28 Instead of storage account sprawl, you'll just have subscription sprawl. Totally not worse.

83:34 Yeah, I think this reduces subscription sprawl, right?

83:36 That's what I said before you added, yeah.

83:38 Oh yeah. I don't, I still never understood what the point of the container is in Azure Blob Store. Like everything's tied to the service account. And so I never really understood.

83:47 It's a prefix.

83:48 It's it? Like that's, that's dumb.

83:49 It's pretty much a prefix.

83:51 Yeah.

83:52 It's no different than AWS, at least as far as I've ever used it. It's no different than a prefix. You can do lifecycle policies, you can do everything at that level. It's just a secondary level down. I mean, that's as far as I've ever really seen it used. I think, I think even like replication and setting your level, if it's GZRS or RZS or any of the other ones, if you do multi-region, like GZRS, you know, is all done at the storage account, not the container level. So I think it's just a prefix, a logical way for a human to segregate it.

84:29 Now you can set Azure Front Door profiles at route-level WAF policies. Coming to Azure Front Door Premium near you, you can now define your WAF policies at the profile and route level, and teams can stop managing a separate policy for every single domain. This will cut operational overhead for organizations running many domains or apps behind one Front Door profile, and a single baseline policy can enforce consistent protection with route-level policies layered on top for apps that need stricter strict your custom rules. Features only available in Azure Front Door Premium. So standard tier customers are out of luck as usual. Premium is billed as base fee plus usage and WAF policy and rule charges will apply separately.

85:10 Another way they charge for security. I mean, the problem is from standard to premium is like a 10x cost. I want to say it's like $35 to $350. But this is a nice feature if you do have a bunch of stuff behind your Route53, behind your same front door, you, you can control stuff a lot better. Before, you either have to have multiple front doors or you just had to do your least privilege or, sorry, most privileged rule set. In that case, that would apply to everything. So again, that granularity is nicer. It just sucks that it's on premium, but We'll bypass Matt's rant about Azure and premium, given that we're like an hour plus into the podcast at this point.

85:56 So see episodes 343, 247.

86:02 I wonder if Bolt could do that. Find me all the episodes when Matt complained about App Gateway and Azure's pricing strategy.

86:10 Another follow-up for next show.

86:13 Damn it, right?

86:14 Haha. Now in Azure, Network Watcher VNet flow logs filtering. So now you can filter your VNet flow logs and you can see only the traffic you care about instead of everything crossing your virtual network. Feature's now generally available and lets you scope the flow logs by IP address, ports, protocols, encryption state, flow state, and traffic direction. So you can only capture the flows that are relevant to you. The main benefit obviously is cost and noise reduction. Flow logs are built on volume of logs collected plus storage, and any kind of log analytics and traffic analytics are on top of that. And so logging less data should be a lower downstream spend. Check the Azure network pricing for current per gigabit rates as cost will vary with usage. This is useful for security and compliance teams who will not choose it because they want to log everything, but they shouldn't, who want to focus on specific signals such as unencrypted traffic, denied flows, and traffic consists of ports. Applies to VM NetFlow logs, the successor to NSG flow logs. So add this use case for migrating off of the NSG flow logs.

87:20 Everybody does want to capture everything till they get that bill, and it is always a shock the first time. It is just— I remember the first time I got a Cisco switch turning on NetFlow and looking at it, I was like, yeah, when they says it captures everything, it's not an exaggeration.

87:36 Yeah, it's everything.

87:39 I actually find that this could be good at doing the opposite, right? You know, you mentioned it there, but a security team just wanting to see what's on unencrypted traffic and then watching that diminish over time. You know, that's an interesting use case given that you're allowed to shrink it down in such a way. So it could be a nice thing. They are clearly deprecating it because they're moving everything away from nginx flows and adding only features to VNet flows, but Azure doesn't deprecate stuff very well. They just slowly make it go into oblivion.

88:11 I mean, it's still problematic, you know, a lot of compliance frameworks say you need to have sort of flow log tracing at this, this level. And, um, it doesn't really dictate the specifics. And so like, you know, if you're savvy, you know that you can filter out to only the traffic you need to see, right? Do you really need to log traffic that's out you know, like, that's outbound from all subnets or, or East-West traffic? Like, probably not. You know, like, that's not the same level of visibility you need for, you know, doing forensic investigation or detecting sort of abnormal usage and access. But, uh, someone does have to sort of define that carefully, because it would suck if you had an exploit that wasn't caught because you filtered away all the signal.

88:52 Yeah, that's always a risk.

88:53 Yep.

88:54 I remember hearing a story, they're like, yeah, we knew this place was compromised for 5 years because we had 5 years of logs. So it was 5+, you know? So it is, yeah, it's, your logs are invaluable in an investigation, but you know, we have not unlocked the infinite money glitch.

89:14 So before we dive into Oracle real fast, I did ask Bolt a few questions as we were doing it, and one of the things he goes, Matt does rant a lot about Azure. And I'm like, thank you, Bolt. I was like, oh.

89:29 Vault knows you so well.

89:31 Right. Thank God we burnt tokens on that.

89:35 Or, bummer, the search results don't surface anything about Matt ranting about the Azure Application Gateway. None of the episodes that came back have listed 312, 320, 321, 329, 337, and others, but don't touch specifically on the App Gateway. And then he goes and lists what those episodes we did read. Ransomed about. So thank you, Vault, for knowing us way too well.

89:56 The Cloud Pod is looking for sponsors to allow us to continue asking, uh, random questions about our own shows, and we need more tokens. So, uh, you know, sign up.

90:07 On to the lovely world of Oracle, who is introducing Console AI, a new simpler way— definitely not accurate— to work with OCI without lawyers involved. Oracle made Console AI generally available, a conversation assistant built into the OCI console. Just what I've always wanted. Console AI lets users describe in plain English what they would like, such as creating a private object storage bucket with versioning and Vault and a Vault key. That's definitely not plain English. Troubleshooting a missing security list or ranking cost drivers directly into the console and having it answer back. It essentially makes OCI Answers equivalent to other assistants that are already in AWS and Azure. Just what I want, another chatbot built into my cloud.

90:53 Okay, but counterargument, have you seen the OCI console? It's awful.

90:58 Counterargument, have you ever seen the IBM console?

91:01 No. Do they have a chatbot? Yeah. I mean, I like some of these things, but I also, yeah, I also think this is kind of a weird use case where it's, I haven't really had that much difficulty navigating consoles, but I'm sure lots of people have.

91:17 I always get so excited 'cause you read like the marketing fluff, you're like, this is gonna be so awesome. And you ask it to do something basic and it's like so off base. Yeah. I don't know if just the context of the platforms are too big or what, but it's like,

91:32 uh, have you ever asked Copilot? Like I've done stuff on where I get like internal server errors. On Azure, and then I tell Copilot to go look at it. He goes, we can't, please go open a support ticket. I'm like, that's not useful.

91:44 Yeah.

91:46 On to other clouds, specifically Cloudflare introducing a full transaction detection for Cloudflare WAF. Cloudflare just added full transaction detection and response rules to the WAF, letting it inspect what the origin returns and not just what the request looks like. Full Transaction Detection correlates each HTTP request with its response to confirm whether the exploit actually succeeded, reducing tradeoffs between broad signatures that cause false positive and narrow rules that miss ambiguous traffic. Rules can be replacing— place a matching rule with a WAF block page or HTTP 403 before it reaches the requester. This prevents sensitive data from getting exposed. But cannot undo. Detection and mitigations are deliberately separate with this current setup, such as detection attached metadata to fields including cf.laf.signatures.response.* and customers decide what to do with them, with the results visible in the security analytics and security events before blocking policy is enabled. The first managed protection target was recently disclosed with PHP to Shell vulnerability, and CloudFront frames the feature as a response to AI agents that adapt payloads based on the response feedback and chain vulnerabilities, or the Cloud Pod trying to get show notes.

93:15 Yeah, I mean, it makes sense, right? The, you know, as someone who had to, uh, recently write some WAF rules for the PHP to Shell vulnerability, it turned out to be like something I thought was something I would be able to, oh, just block this easy thing. Upon research, it was like, oh wait, no, that's not really how it's observed. Um, and it's, it is sort of one of those things where it's like these, these attacks are becoming more and more complex. It's not as simple as just, you know, sort of scripting in the payloads or something that's easy sort of detect with, you know, like a regex type scan. It's, it's more adaptive. And so like the PHP to Shell is exact, it's a good example of that where, you know, what you're inspecting is anything, right? It's the way it's uploading the payload and getting that execution, right? And it's very difficult to sort of detect that just from the HTTP headers. So you want something that's going to scan the file payloads, which is not something normal WAFs do. But now apparently in CloudFront, you can, which is pretty sweet.

94:20 It's just another one of those everything is different now, you know, like just the AI attacks, just traditional security tools and stuff. They're just new tooling that you just wouldn't be standard in any organization now. Um, and this is just one of those things that are just going to be basic necessities.

94:39 It reminds me of just the entire paradigm shift of just moving to the cloud too, where like From the security tool perspective, I think they were running like, I don't know, maybe it was like Qualys or something. And somebody's like, this IP address has a vulnerability. And I was like, that IP address isn't used. They're like, but, but it was, and here it is. I'm like, okay. And then like the next week it was a new one, but it was like a Windows vulnerability and the server was Linux. And I'm like, how do I explain to someone that's saving me a spreadsheet? Like, it's like you, you're speaking different languages, and that's what we're having to come up with, you know, on the fly, essentially, is how do we protect in a completely new paradigm everything that's occurring. And this is just a tool for our tool belt to do it with them.

95:27 Yeah, I mean, think about, you know, traditional WAF rules are like point-in-time inspection of a rule, right? And so, but with AI workloads, you're going to have to have contextual enforcement. That's sort of like, if this request comes in this many times, form this way or irregular in this, you know, then do an action. And it's, that's the, the move that we're going to have to start seeing for all of these things. So like looking for things that are retries and anonymous traffic that way based off of, and then with more contextual data about what it's trying to do and what it's trying to hit. So that's, it's a brand new world. And, uh, man, I was really hoping to be less busy, but It's not gonna happen.

96:06 A whole new world. I won't sing for everyone.

96:09 Yeah.

96:10 Yeah.

96:11 Please don't.

96:12 Title of your next podcast.

96:15 Yeah.

96:17 Well, we made it to the bottom. A little bit of a long show, guys, but we successfully made it there.

96:22 We always do this with such the best, you know, intentions. Like we're cutting, we cut like a third of the stories, maybe even more. And still.

96:29 Yeah. Well, the teacher will be back next week, right?

96:32 Hopefully.

96:32 The problem is with AI, we've, it enables us to add notes and add stories so much easier and they're all interesting, but finding the cream of the crop is hard.

96:43 Yeah.

96:44 Josh, should we scare you off of ever doing this again with us?

96:47 I don't know. Let's see what the, uh, the response from the community is. You know, I'll, I'll go check the, uh, the comments and see if I cry or if I, uh,

96:59 Yeah, YouTube comments and hit us up in Slack. And I recommend that everyone try to make Josh cry.

97:04 Yeah, he's in there. Feel free to add Josh.

97:08 It'll only make me stronger.

97:10 That's the right response.

97:15 All right, guys, it's been a great show. I'll talk to you all later.

97:18 Bye, everybody.

97:20 Another week of cloud news wrapped up. Vault will collect the news, Justin will get the notes, Jonathan will write some code, Ryan will watch the perimeter, and Matt will reluctantly watch Azure. Till next week for AI, Amazon, Google Cloud, and Azure, and hey, maybe even Oracle, who knows? Check out TheCloudPod.net for our newsletter, join our Slack, message us on socials, or leave a review.