OpenAI says its AI agents found 948 Linux kernel flaws for $1.21 million
948 distinct Linux flaws reported in OpenAI test
OpenAI says its autonomous security agents found 948 distinct vulnerabilities in the Linux kernel. The agents also developed working exploits and drafted fixes for review.
The test cost $1,213,430 in API credits. That is about $1,279 per vulnerability.
The Linux kernel is core software used in cloud servers, Android devices, and much critical computing infrastructure.
Key numbers from the October run
- About 2.8 million lines of kernel code were scanned, about 7% of the project.
- 2,308 candidate findings were identified and analysed.
- 1,724 vulnerabilities were dynamically reproduced.
- 948 local privilege-escalation exploits were produced.
- The cost was $1,213,430 in API credits.
What the flaws allow
The flaws are local privilege-escalation vulnerabilities. In simple terms, an attacker first needs limited access to a machine. The flaw could then allow that user to gain root access and take full control.
OpenAI grouped exploits by underlying cause and proposed fix. After grouping, it counted 948 distinct flaws.
OpenAI also said the agents used at least 10 exploitation techniques that were unfamiliar to its researchers and outside experts consulted. It said these were not only existing tricks retrieved from training data.
How this compares with the earlier run
An earlier GPT-5.6 Sol campaign in July found 166 unique privilege-escalation vulnerabilities for $513,690.
A newer, unreleased cyber model found more than five times as many in October. OpenAI said this happened even after months of fixes to the kernel.
OpenAI compared the October cost with historical $20,000 to $70,000 payouts for Linux privilege-escalation exploits at hacking competitions. It suggested a hypothetical value of $18.9 million to $66.3 million if the same results were found manually.
What the agents did not replace
The agents generated candidate patches. Human review and software maintainers remain responsible for validating and implementing fixes, according to OpenAI.
Related reports on AI finding flaws
The source article also describes other recent reports. Last week, Anthropic disclosed that an unreleased version of Claude had discovered and exploited serious vulnerabilities in widely used software, including Chrome and Linux. Some of those exploits could execute malicious code remotely with little or no user interaction.
Researchers at Vidoc Security reproduced several vulnerability discoveries associated with Anthropic's Mythos AI using publicly accessible models for less than $30 per file. The article notes that reproducing a flaw is different from developing a reliable exploit for a protected system.
Why OpenAI says access is restricted
OpenAI warned that attackers with limited access to an outdated operating system may increasingly be able to escalate privileges with AI assistance.
It said it is attempting to restrict access to its more capable cybersecurity models through verified security programs. Its Daybreak Red initiative provides approved defenders with specialised AI tools for vulnerability research, exploit testing, and mitigation development.
Sources
Newisty Editorial Team covers technology, cryptocurrency, digital products, online platforms, developer tools and the wider digital economy. Our content is researched from official sources, company announcements, public documentation, market data and other primary or reputable sources. Articles are reviewed and edited before publication for clarity, accuracy and useful context.