AI Lets Amateurs Find Critical Zero-Days in Hours | Max Justicz

An amateur found a root bug in apt in hours

A researcher who calls himself okay rather than elite aimed GPT-5.6 at apt and found, in a couple of hours, a bug that lets anyone on the network force every Debian and Ubuntu machine to install software as root. Vulnerability discovery has stopped being a specialist skill. Models can search critical code faster than its maintainers can patch it, and the labs now decide who may point them where. Justicz wants defenders to aim the models at their own code before someone else does.

Takeaways

• AI found a critical apt bug in hours. Max Justicz aimed GPT-5.6 at apt code he had audited years earlier and got back, after a couple of hours, a man-in-the-middle bug that makes every Ubuntu and Debian machine install arbitrary software as root.
• Fixes stall on people, not technology. He reported the bug in September and the patch took five or six weeks, with maintainers on vacation for part of that stretch. He says a fix could have shipped the same day.
• Labs are locking down vulnerability hunting access. OpenAI has since restricted Daybreak Blue, its easiest-to-access elevated security model, so his original requests no longer work. He could not get Claude to do anything.
• His edge over ordinary users will shrink. Justicz calls himself okay, not one of the world's best hackers, and says AI gave him a massive uplift. He expects that advantage to shrink until ordinary users can find bugs like his.
• A patched Android phone fell in an afternoon. He bought a modern, fully patched Android phone, aimed the same models at it, and found in one afternoon a way to take it over with a malicious message sent over cell networks. The patch is still rolling out.
• Users can only shrink their exposure. He says ordinary people are in a bad spot, with little to do beyond keeping devices updated and sticking to major services. Markets are not pricing in major attacks, and the tools are not in enough hands yet.

How maintainers should use the models

1. Sign up for OpenAI's Daybreak Blue or Anthropic's cyber verification program for elevated security access.
2. Point the model at your own codebase to audit it. Justicz calls that table stakes now.
3. Spend $100 with Obliteration.ai and aim an obliterated GLM 5.3 at the code as well. It is dumber than GPT-5.6 but good enough for shallower bugs.
4. He sees misaligned models sold as a service as net bad long term but useful now, since defenders can audit faster.