THE CRUNCH
In a single run costing about $1.2 million in API credits, an autonomous pipeline driven by an unreleased OpenAI cyber model produced working root exploits for 948 distinct Linux kernel vulnerabilities, according to the company. The mid-October scan targeted recent code in the kernel's master branch. The agents first filtered the kernel down to the security-relevant code that unprivileged accounts can reach, flagged 2,308 candidate bugs and confirmed 1,724 of them.
Testing took place on virtual machines with an untouched desktop Linux installation, with standard mitigations including KASLR, SMEP and SMAP switched on. In total, the agents produced 1,169 working local privilege escalation exploits, which let a low-level user gain administrator rights, plus 291 proofs of concept for information leaks. OpenAI says the 948 figure reflects separate patches created for the bugs behind those exploits, and that the run cost roughly $1,279 per exploit.
The company claims the agents discovered and used at least 10 novel Linux exploitation techniques that outside kernel experts were not familiar with either, with technical details kept private for safety reasons. An earlier scan in late July yielded 166 unique exploits at a cost of $513,690. OpenAI also notes the October scan found more than five times as many new exploits even though the kernel had gone through three months of hardening in between.
For scale, OpenAI points out that Pwn2Own and TyphoonPWN have historically paid $20,000 to $70,000 for a Linux kernel privilege escalation exploit, so the 948 bugs would fetch $18.9 million to $66.3 million at those rates, between 16 and 55 times the scan's cost. Before anything reached kernel maintainers, OpenAI's security researchers checked the exploits, fixes and proposed severity ratings. The company says hundreds of reports and patches for critical flaws have already gone to the kernel community, with fixes for the remaining bugs to follow in stages through partners in its Patch the Planet initiative. Within 10 days of the scan starting, it observed more than 100 of the flaws patched or carrying staging patches from other contributors.
OpenAI separately raised payouts in its bug bounty programme on November 2, saying the midpoints of its payout ranges for medium, high and critical severity bugs have risen more than fourfold on average. A critical flaw in ChatGPT, Codex or the API Platform can now earn up to $30,000, and bonuses can push a single payout to $172,000.