OpenAI says its autonomous security agents identified 948 distinct vulnerabilities in the Linux kernel, developed working exploits for them and drafted fixes for human review, all for roughly $1.2 million in API credits.
In results released Wednesday, OpenAI said its 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." The October run cost $1,213,430 in API credits, about $1,279 per vulnerability.
The Linux kernel powers cloud servers, Android devices and much of the world's critical computing infrastructure, and 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. "We chose Linux to study how effectively frontier models could find vulnerabilities and demonstrate their security impact in a codebase of this scale and importance," OpenAI said.
OpenAI compared the cost with historical payouts of $20,000 to $70,000 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 campaign using the GPT-5.6 Sol model found 166 unique privilege-escalation vulnerabilities in July for $513,690, while a newer, unreleased cyber model found more than five times as many in October, even after the kernel had undergone months of fixes. OpenAI said its agents used at least 10 exploitation techniques unfamiliar to the researchers and outside experts they consulted, not merely existing tricks retrieved from training data.
The agents scanned about 2.8 million lines of kernel code, roughly 7% of the project. OpenAI said the system 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 attempting to restrict access to its more capable cybersecurity models through verified security programs that give approved defenders specialised tools for vulnerability research, exploit testing and mitigation development. Last week, an unreleased version of Anthropic's Claude was reported to have discovered and exploited serious vulnerabilities in widely used software, including Chrome and Linux, with some exploits capable of executing malicious code remotely.
Analysis
Why This Matters
- Frontier AI systems are now finding real, exploitable vulnerabilities in the world's most widely used operating system kernel, shifting the economics of both cyberattacks and defence.
- Linux underpins cloud servers, Android devices and much of critical infrastructure; privilege-escalation flaws can convert limited access into complete control of a machine.
- OpenAI says it will restrict access to its most capable cybersecurity models, but warns that attackers may increasingly use AI to escalate privileges on outdated systems.
Background
The Linux kernel is the core component of the operating system that runs most cloud servers, Android devices and a large share of the world's critical computing infrastructure. A local privilege-escalation vulnerability lets someone who already has limited access to a machine raise their privileges, potentially to root, and take full control. OpenAI's experiment is part of a wider trend of AI labs using frontier models for cybersecurity research, with both defensive and offensive implications.
What to Watch
- Whether Linux maintainers review and merge the candidate patches OpenAI says its system generated.
- How quickly the disclosed vulnerabilities are fixed, and whether OpenAI expands its verified security program for defenders.
- Whether other AI labs report similar autonomous vulnerability discovery at scale, and how the broader exploit market responds.
Checked
12 claims in this story checked against 1 source: 12 supported.
Show each claim
- SupportedOpenAI’s agents identified 948 distinct vulnerabilities in the Linux kernel.Decrypt: “OpenAI says its autonomous security agents discovered 948 distinct vulnerabilities in the Linux kernel”
- SupportedThe agents developed working exploits for the vulnerabilities.Decrypt: “OpenAI says its autonomous security agents discovered 948 distinct vulnerabilities in the Linux kernel, developed working exploits and drafted fixes for review”
- SupportedThe October run cost $1,213,430 in API credits, roughly $1.2 million.Decrypt: “The October run cost $1,213,430 in API credits”
- SupportedThe flaws could let an unprivileged local user gain root access.Decrypt: “produced 948 local privilege escalation (LPE) exploits that could allow an unprivileged local user to gain root access”
- SupportedOpenAI’s quoted prototype statement says it analyzed 2,308 candidate findings, reproduced 1,724 vulnerabilities and produced 948 LPE exploits.Decrypt: “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”
- SupportedOpenAI said, “We chose Linux to study how effectively frontier models could find vulnerabilities and demonstrate their security impact in a codebase of this scale and importance.”Decrypt: ““We chose Linux to study how effectively frontier models could find vulnerabilities and demonstrate their security impact in a codebase of this scale and importance,” OpenAI said in results Wednesday.”
- SupportedThe run cost about $1,279 per vulnerability.Decrypt: “The October run cost $1,213,430 in API credits, equivalent to around $1,279 per vulnerability.”
- SupportedThe agents scanned about 2.8 million lines of kernel code, roughly 7% of the project.Decrypt: “OpenAI's agents scanned about 2.8 million lines of kernel code, or roughly 7% of the project.”
- SupportedOpenAI compared the results with historical competition payouts of $20,000 to $70,000, suggesting a hypothetical manual value of $18.9 million to $66.3 million.Decrypt: “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.”
- SupportedAn earlier GPT-5.6 Sol campaign found 166 unique privilege-escalation vulnerabilities in July for $513,690.Decrypt: “An earlier GPT-5.6 Sol campaign found 166 unique privilege-escalation vulnerabilities in July for $513,690.”
- SupportedOpenAI said the agents used at least 10 techniques unfamiliar to the researchers and outside experts they consulted.Decrypt: “OpenAI also said its agents used at least 10 exploitation techniques unfamiliar to the researchers and outside experts they consulted”
- SupportedLast week, an unreleased version of Claude was reported to have discovered and exploited serious vulnerabilities in software including Chrome and Linux.Decrypt: “Last week, that an unreleased version of Claude had discovered and exploited serious vulnerabilities in widely used software, including Chrome and Linux.”
