AI has changed everything. Mythos/Cyber has made the Linux virtual machine sandbox useless: https://blog.trailofbits.com/2026/08/26/vms-wont-contain-cyb...

> The only way it might hold here is if no one with access to the cutting edge models bothers to run your code through one while an adversary does. That seems highly unlikely, particularly for any well funded organization.

But since AI models are non-deterministic and the harness matters a lot, simply "running your code through one" doesn't guarantee you'll find all the zero days. An attacker can simply run a few more iterations or use a slightly more advanced harness, or increase the temperature level and they are likely to find zero days which the maintainers didn't.

> But since AI models are non-deterministic and the harness matters a lot, simply "running your code through one" doesn't guarantee you'll find all the zero days.

The exact same line of reasoning applies to human security researchers and pen testers as well. The only difference is that automated testing via frontier models is affordable even for individuals.

The VM comment is total nonsense. That's only going to be the case for an extremely short period here until the major projects all adapt their release procedures to include automated security audits. I imagine the kernel as a whole and KVM in particular will begin receiving a huge amount of such attention. I also expect increasing amounts of such load bearing code to be formally verified as the primary barrier to that has always been the amount of human time required.