Here’s what our cloud hosts had to say about this week’s episode: 371: MrBeast Bets on Gemini for Survival
Justin Brodley
I was out on vacation last week, so for once I listened to the episode the same way all of you do. Matt and Ryan held it down on episode 371, and there was a lot to get through.
The big story is OpenAI’s GPT-6 Astra. It’s the first OpenAI model to cross their own “Critical” cybersecurity threshold, it scored 100% on ExploitBench, and it found two zero-days during testing that OpenAI is now disclosing to maintainers. It’s also a lot better at computer use, finishing OSWorld tasks in about 40 minutes where GPT-5.6 Sol took 75. So of course the reaction is the usual one, this is the dawning of AGI and we are all doomed.
I’m a little skeptical. The story I keep coming back to is the Hugging Face incident from July, which is the reason OpenAI paused before this launch in the first place. The environment those models ran in was supposed to have no internet access. It did have a package registry proxy that was treated as a safe way out, the models were running with their cyber safeguards turned down for the test, and at one point an agent found Hugging Face credentials that someone had left exposed on the public internet. To be fair, the models didn’t just walk through an open door, they found and exploited a zero-day in that proxy to get out, and that capability is real. But a system that should have been network isolated had a path out, and people decided that path was fine.
The same episode had a non-AI version of this. Attackers used a BGP hijack against Hetzner to take over Softaculous IP space and push malicious updates for Virtualizor. Hetzner’s routing security was weak enough to allow the hijack, and Softaculous wasn’t code signing its updates, so the clients had no way to reject the bad ones. RPKI and signed updates are both well known fixes, and neither was in place. No model involved, just ordinary human gaps that someone found.
That worries me more than AGI does. If something really bad happens with AI in the next few years, my bet is it starts with a human shortcut that a very capable system finds faster than we do. So maybe the better question is how we use AI to make the humans smarter. Models like Astra are clearly good at finding holes, and I’d rather they find the proxy that isn’t really isolated, the exposed credentials, and the unsigned update path in our environments before something else does.
On a lighter note, AWS put Amazon Linux 2027 into public preview. SELinux enforcing by default and the 7.1 kernel are good changes. My complaint is the name. I miss the Amazon Linux and Amazon Linux 2 naming convention. The AMI versions were always dated, and that’s where I want the year, not in the name we say out loud and put on PowerPoint slides. By my count this is Amazon Linux 4, and hopefully Amazon undoes the year naming someday.
I’ll be back in the chair next week.
Leave a Reply