OpenAI's autonomous pipeline built 948 Linux root exploits for $1.2M

OpenAI says it has built an autonomous pipeline that hunts for Linux kernel flaws — and the first real run turned up working root exploits for 948 separate vulnerabilities. The system scans code, flags suspects, then tries to break into them in isolated test environments to prove the bug is real. OpenAI announced the pipeline on November 4, 2026, with the first evaluation run carried out in mid-October against the Linux kernel's master branch.

The receipts are blunt. The run burned through $1,213,430 in API costs, which lands at an average of $1,279 for every functioning exploit. The pipeline examined roughly 2.8 million lines of security-relevant code — about 7% of the entire kernel — and flagged 2,308 candidate vulnerabilities from that slice alone. It reproduced 1,724 of them in testing and generated proofs-of-concept for 291 information-leak cases.

The showpiece is the 948 local privilege escalation exploits, or LPEs. An LPE doesn't get an attacker through the front door; it lets someone who already has ordinary access on a machine promote themselves to root, the account that owns everything. According to the research findings, these exploits worked even on standard desktop Linux setups running typical security mitigations.

Fixing all that is another matter. As of the announcement, 61 of the vulnerabilities had been patched upstream, with another 43 moving through a staged patch process. That's 104 bugs fixed or in flight — against 948 working exploits, leaving a hefty backlog still waiting on human review.

This isn't OpenAI's first swing at the kernel. In June 2026 the company published its "Patch the Planet" disclosures, which highlighted 24 LPE exploits surfaced by an earlier model, GPT-5.5 Cyber. Going from 24 to 948 confirmed root exploits in about five months is roughly a 40-fold jump — though OpenAI's own framing notes the comparison isn't perfectly apples to apples, since the new number comes from a dedicated autonomous pipeline rather than a model, and the two efforts may have had different scopes.

For defenders, the number that should sting is that $1,279 average. Kernel-level root exploits have historically demanded skilled specialists and serious time. And since this run covered only about 7% of the kernel, the untouched remainder is presumably not bug-free — future runs could widen the net.