OpenAI's AI built 948 Linux root exploits in one run, for $1,279 each

OpenAI published a report on 4 November 2026 called "A thousand fewer paths to root". In a run in mid-October, an autonomous pipeline built around an upcoming OpenAI cyber model scanned the latest Linux kernel, reproduced the bugs it found, and turned 948 of them into unique, working local privilege escalation exploits. Each one takes a normal user account to full root control on a stock desktop Linux distribution with standard protections such as SMEP, SMAP and KASLR switched on.

The three things you need to know

1. The scale. The pipeline produced 2,308 candidate findings. It reproduced 1,724 of them, wrote 291 proof of concept exploits for information leaks, and ended with 948 unique working root exploits. OpenAI says the agents also came up with 10 exploitation techniques it describes as novel.

2. The price. The whole run cost $1,213,430 in API credits, which works out to about $1,279 per root exploit. Using past payouts from contests like Pwn2Own and TyphoonPWN, OpenAI estimates the same exploits would have been worth roughly $18.9 million to $66.3 million in bug bounties, 16 to 55 times the cost. For comparison, a July run with GPT-5.6 Sol found 166 root exploits for $513,690.

3. The fix. The pipeline also drafted candidate patches. OpenAI says it has already sent hundreds of reports and patches to Linux maintainers, and about ten days in, 61 of the root exploits were fixed upstream with 43 more in staging. The Hacker News reported the same figures.

Why it matters

OpenAI's own warning is blunt: on an unpatched Linux system, a well resourced attacker with any low privileged account should be treated as having root. Finding and weaponizing kernel bugs used to take skilled people weeks. If it now costs around a thousand dollars per exploit, attackers and defenders both get a lot more of them, and the time between a fix landing upstream and attackers using the bug gets shorter.

Dany's take

This is the clearest number yet for what AI does to security economics. The good news is that the same pipeline writes patches, and OpenAI is sending them upstream instead of selling exploits. The bad news is that most servers, routers and cloud images do not update fast. If you run Linux anywhere, patch quickly and add a second layer of isolation, like the sandboxes OpenAI itself recommends, gVisor or Firecracker.