OpenAI says AI pipeline found 948 Linux flaws

OpenAI said on Nov. 4 that it had used an automated Linux kernel security pipeline to identify 948 distinct local privilege-escalation findings after variant analysis, part of a mid-October run it described in a report titled "A thousand fewer paths to root." The company said the pipeline was powered by an unreleased cyber model and that the effort cost about $1.2 million in API usage. SecurityWeek later reported the total as $1,213,430.

According to OpenAI's report, the pipeline moved through multiple stages that included vulnerability discovery, dynamic validation, exploit development and patch generation. That suggests an agentic security workflow rather than a traditional static-analysis tool, with the model narrowing attention to security-relevant areas reachable by low-privileged users before flagging candidate bugs. SecurityWeek said the mid-October scan targeted the most recent code in the Linux kernel master branch.

What did OpenAI say the pipeline found?

OpenAI said the run produced 2,308 candidate Linux-kernel findings, dynamically reproduced 1,724 vulnerabilities and generated 1,169 working local-privilege-escalation exploit results before variant analysis grouped them into 948 distinct underlying findings. The company also said the campaign generated 291 proofs of concept for information-leak vulnerabilities.

SecurityWeek reported that the 948 figure reflects separate vulnerabilities behind the exploit results, not just the raw count of exploit attempts. The outlet also said OpenAI distinguished the 1,169 exploit results from the 948 unique vulnerabilities after variant analysis, which matters for comparing the output to bug bounty totals or other exploit counts.

OpenAI said the pipeline examined the latest code in the Linux kernel master branch in a mid-October scan. Neoteo reported the effort selected 5,034 audit targets across about 2.8 million lines of code, or roughly 7% of the kernel codebase, but the company’s own report is the clearest source for the main counts.

How was the research run?

OpenAI said it reproduced vulnerabilities in isolated virtual machines with KASAN enabled, while exploit testing used virtual machines running an unmodified desktop Linux setup with common hardening features such as KASLR, SMEP and SMAP enabled. Neoteo also reported that exploit-development tests were done with HARDENED_USERCOPY enabled.

That setup is important because it suggests the agents were not testing against a stripped-down lab environment. Instead, the reported exploits were evaluated against a more realistic configuration that still had standard kernel mitigations turned on. SecurityWeek said OpenAI’s agents first narrowed the kernel to security-relevant areas reachable by low-privileged users before flagging candidate bugs.

OpenAI also said the system discovered and used at least 10 novel Linux exploitation techniques, although it withheld the technical details for safety reasons. SecurityWeek reported that outside kernel exploitation experts were not familiar with those techniques.

The reported findings concern local privilege escalation, meaning an attacker already has local access with low privileges and is trying to escalate to root on that machine. They do not describe remote code execution against arbitrary Linux systems.

How much did it cost, and what happens next?

OpenAI said the Linux-kernel run cost about $1.2 million in API credits, and SecurityWeek gave the total as $1,213,430. SecurityWeek also reported the run cost about $1,279 spread across the 948 distinct findings.

OpenAI compared the 948 Linux privilege-escalation bugs with historical Pwn2Own and TyphoonPWN payouts of $20,000 and $70,000 each, and said the October scan found more than five times as many new local privilege escalations despite three months of intervening kernel hardening. SecurityWeek also reported that OpenAI had previously run a late-July scan powered by GPT-5.6 Sol that found 166 unique local privilege escalations for $513,690.

OpenAI said it had already submitted hundreds of reports and patches for critical flaws. SecurityWeek reported that it planned to send fixes for the remaining bugs in stages with partners in its Patch the Planet initiative. Neoteo said that at the end of the scan-and-analysis period, 61 local privilege escalations had been patched upstream, 43 had staging patches and 299 were being publicly reported and discussed, while OpenAI did not say all 948 distinct findings had been fixed.

SecurityWeek said OpenAI recommended layered defenses, including sandboxing and isolation tools such as Firecracker and gVisor, to reduce damage if similar flaws are exploited. OpenAI also warned that for the near and medium term, low-privileged access on an unpatched operating system should be treated as functionally equivalent to root access for a well-resourced threat actor.