OpenAI's $1.2M Run Mined 948 Linux Kernel Root Exploits at $1,279 Each

The public price tag of $1,279 per local privilege escalation resets threat modeling for any ops team: unprivileged accounts on an unpatched OS now count as root in front of a well-resourced attacker, and old code must be re-scanned with each model generation. With 948 candidate patches still awaiting human review, kernel maintainers remain the bottleneck, and the fix side is structurally slower than discovery.

OpenAI has run an autonomous vulnerability pipeline, driven by its unreleased cyber model, that scanned roughly 2.8 million lines of security-relevant code from the Linux kernel master branch. The run dynamically reproduced 1,724 vulnerabilities, generated 1,169 usable privilege-escalation exploits plus 291 information-disclosure PoCs, and after variant-analysis deduplication settled on 948 distinct local privilege escalation (LPE) vulnerabilities. OpenAI also produced 948 candidate patches for maintainers. The API bill: $1,213,430, or about $1,279 per root-capable vulnerability.

The findings reset the economics of both attack and defense. A low-privilege account on an unpached OS should now be treated as equivalent to root in front of a well-resourced attacker, OpenAI writes, and the company recommends sandboxing with Firecracker or gVisor, re-scanning old code with each new model generation, and automating patch generation.

The fix side lags. A PHP memory-corruption bug sat unfixed since 2024; an OpenAI-sponsored audit turned it into arbitrary code execution in two AI-assisted weeks, while four engineers spent two months patching it. Columbia's Suman Jana calls the reasoning 'painful and tedious' and well suited to AI. Anthropic's formal verification work on LLM-generated patches points the way to closing that gap, but Johns Hopkins' Matthew Green cautions that real investment will come only after a widely felt incident, advising users to install updates now.