OpenAI's autonomous cybersecurity system found 948 distinct privilege-escalation flaws in the Linux kernel and built working exploits for most of them. The research cost $1.21 million (USD) in API credits. It matters because the same AI tools that find bugs for defenders could soon help attackers do it faster too.
What actually happened
OpenAI published the results on Wednesday, saying its autonomous agents scanned about 2.8 million lines of Linux kernel code, roughly 7% of the project. The agents identified 2,308 candidate findings, reproduced 1,724 vulnerabilities, and produced 948 local privilege-escalation exploits, according to Decrypt. The October run cost $1,213,430 in API credits, about $1,279 per vulnerability. OpenAI compared that with $20,000 to $70,000 payouts for similar exploits at hacking competitions, suggesting a manual-discovery value of $18.9 million to $66.3 million. An earlier GPT-5.6 Sol run found 166 unique flaws in July for $513,690. A newer, unreleased model found more than five times as many in October. OpenAI said agents used at least 10 exploitation techniques unfamiliar to its own researchers and outside experts.
How we got here
Linux powers cloud servers, Android devices, and much of global computing infrastructure. OpenAI said it chose Linux to study how effectively frontier models could find vulnerabilities and demonstrate their security impact in a codebase of this scale and importance, per Decrypt. Last week, Anthropic disclosed that an unreleased Claude version had found and exploited serious vulnerabilities in Chrome and Linux, some allowing remote code execution with little user interaction. Separately, researchers at Vidoc Security reproduced vulnerability discoveries tied to Anthropic's Mythos AI using public models for under $30 per file. Together, these disclosures show AI-driven vulnerability research moving from expensive, manual work toward automated, lower-cost systems within months.
Why this matters for you
For software maintainers, this means patch review will matter more than discovery. OpenAI's system drafted candidate fixes, but the company said human reviewers and maintainers remain responsible for validating and implementing them. For builders of connected hardware, including wearables and smart glasses that run Linux-based systems, faster vulnerability discovery raises the stakes for timely updates. For everyday users, the near-term risk stays theoretical, since these exploits required local access first. OpenAI is restricting its most capable tools to vetted defenders through its Daybreak Red program, but it warned that attackers with limited access to outdated systems may increasingly escalate privileges with AI help.
The bigger question
If AI systems can find and exploit hundreds of flaws in mature, heavily audited code for roughly $1,279 each, how long can defenders stay ahead of attackers using the same tools? OpenAI frames its results as a win for defense, since patches can ship before criminals find the same bugs. But the same automation could flow to less scrupulous actors as capable models grow cheaper. Who controls access to these tools may decide which side benefits most.
What to watch
OpenAI published its Linux results on Wednesday, without naming a public release date for the newer cyber model used in October. Its Daybreak Red program continues onboarding vetted defenders for vulnerability research and mitigation work. Watch for Linux kernel maintainers' response to the disclosed flaws, and for whether Anthropic or other labs publish comparable exploit counts.
