OpenAI’s Linux cybersecurity pipeline
OpenAI described an AI-agent pipeline that used an upcoming, cyber-focused model to examine Linux kernel security issues. In a run in mid-October 2026, the system dynamically reproduced 1,724 vulnerabilities and generated 948 local privilege-escalation (LPE) exploits. The run took 30 hours and cost $1,213,430 in API credits, according to the company.
The prototype identified and analyzed 2,308 candidate findings in the Linux kernel. It also produced proofs of concept for 291 information-disclosure vulnerabilities. OpenAI says the 948 LPE exploits could let an unprivileged local user gain root access. Those figures describe the results OpenAI reported for that run; they do not establish that every exploit was independently verified or that the pipeline will achieve the same results on other code or systems.
How the five-stage process works
The pipeline is organized into five stages, from finding candidate weaknesses to human review. In the discovery stage, agents select high-value areas of code to audit and identify potential findings, focusing on security-relevant parts that low-privilege users can reach.
During dynamic validation, agents try to reproduce each candidate in a sandbox and preserve runtime evidence. The exploitability stage tests whether a reproduced weakness could enable information disclosure or privilege escalation, and produces a proof-of-concept exploit. Agents then generate proposed patches. In the final stage, security researchers review the exploit, patch and proposed severity before the finding is reported to maintainers.
The five stages in OpenAI’s cybersecurity pipeline
OpenAI’s reported workflow from candidate discovery to human review.
| Stage | Main function | Reported output |
|---|---|---|
| Discovery | Identify audit targets and candidate findings | 2,308 candidates analyzed |
| Dynamic validation | Reproduce candidates in a sandbox and preserve runtime evidence | 1,724 vulnerabilities reproduced |
| Exploitability assessment | Test potential abuse and develop proofs of concept | 948 LPE exploits and 291 information-disclosure proofs of concept |
| Patch generation | Propose fixes for identified vulnerabilities | Proposed patches |
| Human review | Experts assess the exploit, patch and proposed severity before reporting | Report to maintainers |
What the reported cost does—and does not—show
OpenAI reported a total cost of $1,213,430 in API credits, or about $1,279 per generated LPE exploit. That figure covers the API credits for the reported run. It is not, by itself, a comparison with the full cost of an equivalent human-led audit, nor does it account for every possible expense involved in validating, reviewing and maintaining fixes.
The company said that, with iteration guided by human experts, AI-generated patches can be functionally correct, fit existing coding styles and withstand attack variants. This is OpenAI’s assessment, not a guarantee that every proposed patch is correct or secure. The pipeline includes human review before reporting, and the reported account does not establish that all patches were accepted or deployed.
What the Linux findings mean for defenders
OpenAI said the pipeline builds on 24 LPE exploits that its GPT-5.5-Cyber model had found in June 2026. The later run reportedly used at least 10 novel Linux exploitation techniques. OpenAI described two new persistent elastic objects for spraying objects in a target kernel cache, two page-grooming techniques intended to improve exploit reliability, five data-only techniques for privilege escalation, and one technique involving arbitrary type confusion.
The results underline why defenders need to assess and patch vulnerable systems promptly, but the reported experiment does not show that every Linux system is exposed to the same weaknesses. OpenAI warned that, in the near and medium term, a well-resourced threat actor’s access to a low-privilege account on an operating system missing its latest available patches should be treated as functionally equivalent to root access. That is the company’s risk assessment, not a universal statement that every unpatched machine can be compromised in the same way.
AI-assisted vulnerability discovery and patching
The pipeline is part of a broader shift in cybersecurity: AI systems can help identify software weaknesses, assess whether they are exploitable and propose fixes. A VulnCheck report described AI-assisted vulnerability discovery as a growing risk, including because it can lower the technical barrier for novice hackers. The same capabilities can also support defenders, but they do not remove the need to verify findings and test patches.
Google Threat Intelligence Group reported that monthly vulnerability disclosures rose from 5,045 in January 2026 to 10,740 in August 2026. Those counts measure disclosures, not vulnerabilities discovered by AI alone. The UK National Cyber Security Centre said defensive AI tools can shorten the time between a zero-day vulnerability and its disclosure as a known vulnerability, and speed up patching. It also cautioned that systems can remain exposed when organizations do not configure and use these tools correctly.
OpenAI argued that automated patch generation needs to keep pace with vulnerability discovery to reduce the period in which software remains exposed. Its reported pipeline is an example of how AI may contribute to both sides of that work: generating exploits that demonstrate potential impact, and proposing patches for human assessment. The demonstrated figures apply to one reported run; they do not establish a general rate of discovery, a guaranteed reduction in risk or a substitute for independent validation and responsible disclosure.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction