OpenAI has published results from an autonomous pipeline, run on a cyber model it has not released, that hunted for local privilege-escalation bugs in the Linux kernel, exploited them and drafted patches. The scan ran in mid-October against the kernel's master branch, with agents narrowing in on security-relevant code that a low-privileged user can reach. Researchers checked each exploit, fix and proposed severity before anything went to kernel maintainers.
“We chose Linux to study how effectively frontier models could find vulnerabilities and demonstrate their security impact in a codebase of this scale and importance”
OpenAI, in its published results, in its published results · decrypt.co ↗
The numbers are large. From 2,308 candidate findings, 1,724 were reproduced, and the work came down to 948 distinct bugs, counted as the separate patches behind the exploits. The run cost about $1.21 million in API credits, roughly $1,279 a bug, and covered about 2.8 million lines of kernel code, around 7% of the project. OpenAI says at least ten of the exploitation techniques were new to its own researchers and to outside experts, and it is withholding the details for safety.
OpenAI weighs that spend against what such bugs fetch at hacking contests like Pwn2Own and TyphoonPWN, where payouts of $20,000 to $70,000 imply a market value of $18.9 million to $66.3 million, 16 to 55 times the cost.
An earlier scan in late July, with GPT-5.6 Sol, found 166 unique bugs for $513,690, so October's run found over five times as many despite three months of kernel hardening. Decrypt adds that Anthropic disclosed the week before that an unreleased Claude version had exploited flaws in Chrome and Linux,.
OpenAI's advice to defenders is blunt: for now, an operating system that is not fully patched should be treated as giving a well-resourced attacker root once they hold any low-privilege account. Decrypt notes that automating exploitation may let defenders patch ahead of criminals, though reproducing a flaw is not the same as building a reliable exploit. OpenAI has sent hundreds of reports and patches and will send the rest in stages with partners in its Patch the Planet effort, while stronger cyber models stay limited to verified programmes.
Sources · 2