OpenAI said an experimental AI pipeline generated 948 working Linux exploits that could let a user with limited local access gain root, the system’s highest level of control. The company disclosed the research on Wednesday, The Hacker News reported, describing a process that combines automated vulnerability discovery, exploit testing and patch generation with human review.
The findings concern the Linux kernel, the core software that manages hardware and system resources. OpenAI said the exploits worked against a commonly used Linux distribution. The local-access requirement matters: these are paths from an existing low-privileged foothold to greater control, rather than a reported means of initially entering a machine over the internet.
From candidate flaws to working exploits
In a run conducted in mid-October 2026, OpenAI said its prototype identified and analyzed 2,308 candidate kernel findings and dynamically reproduced 1,724 vulnerabilities within 30 hours. It generated proof-of-concept programs—code demonstrating exploitation—for 291 information-leak vulnerabilities, alongside the 948 local privilege escalation exploits.
The run used an upcoming cyber-focused model and cost $1,213,430 in API credits, which OpenAI put at $1,279 per local privilege escalation exploit. Those figures describe the company’s experimental results and API spending, rather than independently established performance or a complete accounting of research costs.
The work expands on 24 local privilege escalation exploits that OpenAI disclosed in June 2026 as discoveries by GPT‑5.5‑Cyber.
The newer pipeline has five stages. First, AI agents select security-relevant code that low-privileged users can reach and search it for potential flaws. They then try to reproduce each finding in an isolated test environment, retaining evidence of what happened during execution.
Next, the agents assess whether a reproduced flaw can expose information or raise privileges and develop demonstration exploits. A patch-generation stage produces proposed fixes. Finally, human security researchers review the exploit, patch and proposed severity before reporting the finding to maintainers.
OpenAI also said the agents discovered and used at least 10 Linux exploitation techniques that, to its knowledge, were novel. These comprised two new persistent “elastic objects” for placing objects into targeted kernel memory caches; two page-grooming techniques, which arrange memory to improve exploit reliability; five data-only techniques that gain privileges directly or through an arbitrary write, meaning an attacker-controlled memory write; and one technique involving arbitrary type confusion, where software treats an object as the wrong type.
Faster discovery puts pressure on patching
The disclosure comes amid warnings that AI is accelerating vulnerability research for both defenders and attackers. The Hacker News cited an October VulnCheck report describing AI-assisted discovery as a growing risk category that lowers the technical expertise novices need to find software flaws.
The publication also cited Google Threat Intelligence Group figures showing monthly vulnerability disclosures rising from 5,045 in January 2026 to 10,740 in August. Disclosures classified as high-risk rose from 131 to 350 over the same period, an increase of 167%.
In an assessment published Tuesday, the UK National Cyber Security Centre said widespread access to AI-powered cyber tools had “almost certainly” significantly increased discovery of zero-day vulnerabilities—flaws unknown before their discovery. As quoted by The Hacker News, the agency said defensive tools shorten the transition from a zero-day to a publicly known vulnerability and speed patching, but systems remain exposed when the technology is not correctly and widely configured.
The NCSC also assessed that cheaper replacement access has made attackers less cautious about losing a foothold while pursuing an objective, enabling broader use of cyber capabilities to obtain and use information.
OpenAI argued that automated patch generation must keep pace with discovery to narrow the opportunity for exploitation. It said human expert-guided iteration can make AI-generated patches functionally correct, consistent with existing coding styles and resistant to attack variants.
For administrators, OpenAI’s warning was specific: it believes well-resourced attackers should be treated as effectively having root access when they already hold a low-privileged account on an operating system missing its latest available patches. Organizations responsible for patching and testing should adapt those processes to the faster pace of discovery, as the NCSC recommends.
