Key facts
- OpenAI said its agents produced 948 local privilege-escalation exploits, grouped into 948 distinct flaws by cause and fix.
- OpenAI said the agents used at least 10 exploitation techniques unfamiliar to the researchers and outside experts consulted.
- OpenAI said the October run cost $1,213,430 in API credits, about $1,279 per vulnerability.
- OpenAI compared the cost with $20,000 to $70,000 historical payouts for Linux privilege-escalation exploits at hacking competitions.
- OpenAI said a newer, unreleased cyber model found more than five times as many flaws as the July campaign.
- OpenAI said candidate patches still require human review and validation by software maintainers.
OpenAI said its autonomous security agents discovered 948 distinct vulnerabilities in the Linux kernel, developed working exploits and drafted fixes for review, at a cost of roughly $1.2 million in API credits. The company published the results on Wednesday, November 4, 2026.
The flaws are local privilege-escalation vulnerabilities: an attacker would first need limited access to a vulnerable machine, but could then potentially obtain root privileges and take full control. OpenAI said its agents scanned about 2.8 million lines of kernel code, roughly 7% of the project.
"Our prototype identified and analyzed 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 vulnerabilities, and produced 948 local privilege escalation (LPE) exploits that could allow an unprivileged local user to gain root access," OpenAI said. After grouping exploits by their underlying cause and proposed fix, researchers counted 948 distinct flaws.
The October run cost $1,213,430 in API credits, equivalent to around $1,279 per vulnerability. OpenAI compared that with historical $20,000-to-$70,000 payouts for Linux privilege-escalation exploits at hacking competitions, suggesting a hypothetical value of $18.9 million to $66.3 million if the same results were found manually.
An earlier GPT-5.6 Sol campaign found 166 unique privilege-escalation vulnerabilities in July for $513,690. OpenAI said a newer, unreleased cyber model found more than five times as many in October, even after the kernel had undergone months of fixes. OpenAI also said its agents used at least 10 exploitation techniques unfamiliar to the researchers and outside experts they consulted.
OpenAI said the system also generated candidate patches, but human review and software maintainers remain responsible for validating and implementing fixes. The company warned that attackers with limited access to an outdated operating system may increasingly be able to escalate their privileges with AI assistance, and said it is restricting access to its more capable cybersecurity models through verified security programs, including its Daybreak Red initiative for approved defenders.
The report also noted that Anthropic disclosed last week that an unreleased version of Claude had discovered and exploited serious vulnerabilities in widely used software, including Chrome and Linux, and that researchers at Vidoc Security reproduced several vulnerability discoveries associated with Anthropic's Mythos AI using publicly accessible models for less than $30 per file.