OpenAI detailed a prototype pipeline that used an upcoming cyber-focused artificial intelligence model to validate and patch 948 local privilege escalation vulnerabilities in Linux. The company said the run cost $1.2 million in API costs and that the work produced working exploits that could let a low-privileged local user reach root access.
The pipeline and its five stages
OpenAI described the process as an agentic pipeline with five stages. It starts with vulnerability discovery, where agents look for high-value audit targets and candidate findings in code that is security-relevant and reachable by low-privileged users. Dynamic validation follows, with agents trying to reproduce each candidate in a sandboxed environment and keeping runtime evidence. Exploitability assessment comes next, where agents test whether the validated flaws could support information disclosure or local privilege escalation and then build a proof-of-concept exploit. Patch generation then produces fixes for the identified issue. The final stage is human review, where security researchers review the exploit, the patch, and the proposed severity before the issue is reported to maintainers.
What the run produced
In a run in mid-October 2026, the pipeline reproduced 1,724 vulnerabilities within 30 hours and generated proof-of-concepts for 291 information leak vulnerabilities and 948 working local privilege escalation exploits for a commonly used Linux distribution. OpenAI said the total run cost $1,213,430 in API credits, which works out to $1,279 per local privilege escalation exploit. The company also said that, to the best of its knowledge, the agents discovered and used at least 10 novel Linux exploitation techniques.
OpenAI broke those techniques into four groups. Two were new persistent elastic objects used to spray objects inside the target kernel cache. Two were new page grooming techniques meant to improve exploit reliability. Five were new data-only techniques that achieved privilege escalation directly or through an arbitrary write. One technique involved arbitrary type confusion.
What the findings suggest
OpenAI said the result shows how frontier AI is changing cybersecurity by speeding up vulnerability discovery and exploitation while also helping defenders address defects through extra instructions and iterations before attackers can use them. The company said that can make it cost prohibitive for bad actors to exploit software. It also said that, for the near and medium term, having a low-privileged account on an operating system that is not up to date on its latest available patches should be treated as functionally equivalent to root access for a well-resourced threat actor.
Broader warnings from security groups
VulnCheck described AI-assisted vulnerability discovery as a growing risk category in a report published last month. It said the technology can speed up novice hackers, who now need little to no technical expertise to find software vulnerabilities. Google Threat Intelligence Group said the number of vulnerabilities disclosed per month rose from 5,045 in January 2026 to 10,740 in August 2026. It also said high-risk vulnerabilities increased from 131 disclosures in January 2026 to 350 in August 2026, a 167% rise.
The U.K. National Cyber Security Centre said in an assessment published Tuesday that the widespread availability of AI-powered cyber tools has almost certainly led to a significant increase in zero-day discovery. It added that defensive AI cyber tools have shortened the time between zero-day and n-day status and made patching faster. The same assessment said many systems remain vulnerable if the technology is not configured correctly and widely, and that organizations will need to change internal patching and testing processes to keep pace. It also said attackers have shown less caution about losing access when the cost of replacement is minimal, and that cyber capabilities are now used across a broader range of activity because the lower risk makes it easier for adversaries to access information and use it operationally.
Licensed under CC BY-ND 4.0 International