OpenAI has demonstrated how autonomous AI agents can automate large-scale cybersecurity research, from identifying software vulnerabilities to developing working exploits and generating security patches. In an experiment targeting the Linux kernel, an AI-powered pipeline identified 2,308 potential security issues, reproduced 1,724 vulnerabilities, and generated 948 unique local privilege escalation exploits.
The Linux kernel was selected as the target because it underpins Android devices, cloud infrastructure, supercomputers, and numerous other systems. With more than 40 million lines of code, it represents a substantial challenge for traditional security auditing.
Instead of analyzing the entire codebase, OpenAI developed a multi-agent system that focuses on security-relevant components accessible to users with limited privileges. The initial stage identified 5,034 audit targets covering approximately 2.8 million lines of code, or around 7% of the kernel.
The system then coordinated specialized AI agents through five stages:
- Vulnerability discovery: Scout agents identify security-critical areas of the codebase, while Analyst agents examine the selected components for potential weaknesses.
- Dynamic validation: AI agents attempt to reproduce suspected vulnerabilities in isolated virtual machines, using runtime evidence to distinguish actual security flaws from false positives.
- Exploitability assessment: Specialized agents develop proof-of-concept (PoC) exploits to determine whether confirmed vulnerabilities could enable information disclosure or local privilege escalation.
- Patch generation: AI generates candidate fixes, checks them against coding practices, and analyzes potential exploit variations to improve patch completeness.
- Human review: Security researchers examine the findings, exploit demonstrations, and proposed patches before reporting them to Linux maintainers.
This approach combines automated code analysis with practical exploitation testing, allowing the system to assess not only whether a vulnerability exists but also how it could affect system security.
The research also revealed at least 10 previously undocumented Linux exploitation techniques, demonstrating that AI agents can discover new approaches to bypass security protections rather than simply reproduce known attack patterns.
To test the exploits under realistic conditions, OpenAI used virtual machines running a standard Linux distribution with common security protections enabled, including Kernel Address Space Layout Randomization (KASLR), Supervisor Mode Execution Prevention (SMEP), and Supervisor Mode Access Prevention (SMAP).
The experiment cost approximately $1.21 million in API credits, averaging $1,279 per unique privilege escalation vulnerability. For comparison, historical bug-bounty payouts for similar Linux kernel privilege-escalation bugs ranged from $20,000 to $70,000 each.
The results represent a significant improvement over an earlier experiment conducted in July 2026, when GPT-5.6 Sol identified 166 unique Linux privilege escalation vulnerabilities. The newer cybersecurity model discovered more than five times as many, despite additional security hardening between the experiments.
However, increasingly capable AI systems also introduce new security challenges. The same technologies that help defenders identify and fix vulnerabilities could potentially enable attackers to discover and exploit software weaknesses more efficiently.
OpenAI emphasizes that automated vulnerability discovery must be accompanied by faster patching and independent verification. The company plans to extend this approach to other critical software projects, helping developers identify security flaws and address them before they can be exploited.
