OpenAI says an automated vulnerability-research pipeline built working exploits for 948 distinct Linux local privilege-escalation flaws during a mid-October run, demonstrating how frontier AI systems could accelerate both offensive research and patch development. The company said the experiment cost $1,213,430 in API credits and used an upcoming cyber model that has not been publicly identified. OpenAI’s account of the research says the system produced a candidate patch for each of the 948 vulnerabilities.
The pipeline first selected 5,034 security-relevant audit targets covering about 2.8 million lines of kernel code, or roughly 7% of the Linux codebase. Analyst agents generated 2,308 deduplicated candidate findings within 12 hours. Triage agents then attempted to reproduce them in isolated virtual machines, dynamically confirming 1,724 vulnerabilities—about 74.7% of the candidates—within 30 hours.
OpenAI subsequently screened those findings for higher-impact outcomes and produced 291 information-leak proofs of concept and 1,169 working local privilege-escalation exploits. Variant analysis reduced the latter figure to 948 distinct vulnerabilities, each capable, in the tested environment, of allowing an unprivileged local user to obtain root access. The exploit-development agents operated with common Linux defenses enabled, including SMEP, SMAP, KASLR and HARDENED_USERCOPY, rather than relying on a deliberately weakened target.
The results do not yet map cleanly to a list of public CVE records. OpenAI’s disclosure does not provide CVE identifiers, CVSS scores or a claim that these flaws have been exploited in the wild. It says that, after 10 days of autonomous work and manual analysis, 61 of the LPEs had been patched upstream, 43 had staging patches and 299 were being reported or discussed publicly. Administrators should therefore treat the announcement as a warning about exposure and patch-tracking workload, not as evidence that 948 immediately exploitable CVEs have entered production threat activity.
OpenAI said hundreds of reports and patches have already been sent to Linux maintainers and partners, but also acknowledged that automated exploitation and patching require independent review. In an earlier run, researchers found that 20 purported successful escalations had reached root through a different zero-day than the vulnerability assigned to the agents; the pipeline was subsequently modified to verify the exploit path. Linux users should continue applying vendor and distribution kernel updates, monitor security advisories for newly assigned identifiers, and avoid deploying AI-generated patches without testing their correctness, compatibility and resistance to related attack variants.
