OpenAI's AI Built 948 Linux Root Exploits in One Run

OpenAI says an autonomous AI pipeline built 948 working local privilege escalation exploits for the Linux kernel in a single run, then generated candidate patches for the bugs behind them. The company described the work in a research post on Wednesday, Nov. 4.

A local privilege escalation, or LPE, is a flaw that lets an ordinary, unprivileged user on a machine gain root access. OpenAI says its test machines ran a stock desktop Linux system, not a deliberately weakened target.

How the pipeline works

The run took place in mid-October 2026 and used an upcoming OpenAI cyber model. It had five stages:

• Discovery: Scout agents picked 5,034 audit targets, about 2.8 million lines of code, or roughly 7% of the kernel. Analyst agents then produced 2,308 candidate findings.
• Validation: Triage agents reproduced 1,724 vulnerabilities in isolated virtual machines, about three quarters of the candidates.
• Exploitability: Exploit agents built working exploits with standard protections like KASLR, SMEP, and SMAP turned on. After variant analysis, OpenAI counts 948 unique root exploits.
• Patching: The system wrote a candidate patch for each vulnerability.
• Human review: Security researchers check the exploit, patch, and severity before reporting to maintainers.

What it cost

The whole run cost $1,213,430 in API credits, or about $1,279 per root exploit. OpenAI notes that historical Pwn2Own and TyphoonPWN bounties of $20,000 and $70,000 for Linux kernel root exploits would put a hypothetical value of $18.9 million to $66.3 million on the batch.

Stronger models find more

A late-July run with GPT-5.6 Sol found and exploited 166 unique root vulnerabilities. The newer model found more than five times as many in October, despite three months of kernel hardening in between. OpenAI also says the agents used at least 10 exploitation techniques that outside kernel experts hadn’t seen before, and it is withholding their details for safety.

Getting the fixes out

Following guidance from the Linux kernel community, OpenAI says it has already sent hundreds of reports and patches for critical vulnerabilities, and is working with its Patch the Planet partners on the rest. By the time the scan and analysis wrapped after 10 days, 61 of the root bugs had already been patched upstream, 43 had staging patches, and 299 were being discussed publicly.

OpenAI’s takeaway for everyone else: keep systems patched, because on an out-of-date machine, a low-privileged account should be treated as nearly equal to root for a well-resourced attacker.