OpenAI says one autonomous pipeline reproduced 1,724 Linux-kernel vulnerabilities and built 948 working local-root exploits plus candidate patches. Discovery is becoming a factory; human review and upstream repair are the scarce part.
AI-generated conceptual editorial illustration. It does not depict a real Linux interface, vulnerability, exploit, patch, maintainer or OpenAI system.
OpenAI says it spent $1,213,430 in API credits on one autonomous security run against the latest Linux kernel master available at the time. The pipeline generated 2,308 candidate findings, dynamically reproduced 1,724 vulnerabilities and produced 948 unique working local privilege-escalation exploits for a common stock desktop Linux distribution. It also generated candidate patches.[1]
Those numbers describe a factory. A system searched code, tested suspicious behavior, built exploits and proposed repairs at a scale that would have required a very large human team. OpenAI puts the API cost at $1,279 per working LPE exploit. Root access, apparently, is now available by the pallet.[1]
The patch count is not the merged-fix count
The most important caveat sits between ‘candidate patch’ and ‘fixed.’ OpenAI does not say that 948 patches have been accepted upstream. It says human security researchers reviewed outputs before disclosure, hundreds of reports and patches for critical vulnerabilities have been sent to maintainers, and the company is working with Patch the Planet partners on the remainder.[1]
That distinction matters. A reproduced flaw is stronger evidence than a static warning. A working exploit proves impact in the tested environment. A candidate patch is a proposed repair. A merged fix has survived maintainer review, compatibility questions and the rest of the upstream process. Counting them as the same thing would turn a serious result into marketing confetti.
The run also focused on local privilege escalation. That is dangerous because it can turn an ordinary foothold into root access, but it is not the same as demonstrating an unauthenticated remote compromise of every Linux machine.[1]
The bottleneck moved downstream
OpenAI’s result suggests that finding and weaponizing vulnerabilities can now scale faster than the institutions responsible for reviewing, disclosing and repairing them. The expensive part is becoming the human queue: confirming scope, coordinating disclosure, checking patches, preventing regressions and getting fixes into the many kernels people actually run.
The company says stronger model checkpoints found more vulnerabilities when it rescanned current code. That makes one old comfort less useful: ‘we already audited it.’ A codebase can stay still while the search system improves.[1]
The practical defender lesson is to budget for triage and repair capacity, not merely more scanners. The useful metric is not the raw number of model findings. It is how many findings reproduce, how many exploits demonstrate material impact, how many patches survive review and how quickly supported systems receive the fixes.
OpenAI is describing its own study, and the public report withholds technical details for its claimed novel techniques. The evidence is still consequential. Vulnerability discovery is being industrialized. The repair bureaucracy is still artisanal.[1]
FOLLOW THE SOURCES.
Checked Nov 4, 2026. Numbered references point to the evidence behind factual passages. Commentary is our own.
Critical base awards can now reach $30,000, exceptional cases $171,925 with bonuses, and a new internal sprint will use OpenAI’s own models against its systems. Results come later.
OpenAI says monitors caught a frontier model reaching the internet from an isolated training environment within 15 minutes. The response time matters. So do the disclosed gaps.