Some people find one kernel bug and retire on the bounty. OpenAI's machines found 948 of them in a single run, and nobody retired.
On Wednesday OpenAI published research on an agentic pipeline built on an upcoming, unreleased cyber model. In a run in mid-October against the latest Linux kernel, on a stock desktop install of a widely used distribution, the agents produced 948 working local privilege escalation exploits. That's the kind of bug that lets an ordinary logged-in user become root. The total bill was $1,213,430 in API credits, or $1,279 per exploit.
I respect a man who shows his receipts.
How you get to 948
The pipeline has five stages, and only the last one is human:
- Pick targets. Agents narrowed the kernel's 40-million-plus lines down to roughly 7% that low-privileged users can reach and that matter for security.
- Find bugs. A second group returned 2,308 candidate findings in 12 hours.
- Prove them. Triage agents reproduced 1,724 of those in isolated virtual machines within 30 hours, a 74.7% hit rate.
- Weaponize them. Exploit agents produced 291 information-leak proofs of concept and 1,169 working root exploits, which collapsed to 948 distinct vulnerabilities once duplicates were grouped.
- Human review. Security researchers check every exploit, patch and severity rating before anything goes to the kernel maintainers.
The agents also wrote a patch for each underlying flaw. In a sample of 166 exploits, OpenAI says they used at least 10 exploitation techniques that outside kernel experts hadn't seen before. OpenAI kept the details out of the post for safety reasons. Good call.
Pwn2Own has paid $20,000 for one Linux kernel root exploit. OpenAI's own math puts this run's haul at $18.9 million to $66.3 million.
The machine lied about its wins
This is the part that matters. In an early run, the exploitation stage claimed 200 successful root exploits. Human review found that 20 of them reached root through a different zero-day than the one the agent had been assigned. The agent got the right result for the wrong reason and reported it as success.
OpenAI's own takeaway: "even successful exploitation by agents needs proper validation." It tightened the task rules and added an independent check that confirms each exploit actually uses the bug it was meant to use.
Chuck Norris doesn't need a second opinion. Everybody else does.
Patch now
When the scan and manual analysis wrapped after 10 days, 61 of the bugs were patched upstream, 43 had staging patches and 299 were being reported and discussed publicly. The rest are going out through OpenAI's Patch the Planet partners. OpenAI's warning is blunt: for now, a low-privileged account on an unpatched system "should be considered functionally equivalent to root access for a well-resourced threat actor." Separately, the UK's National Cyber Security Centre said this week that widely available AI cyber tools have "almost certainly" driven a big rise in zero-day discovery.
So update your machines. Today.
The lesson for everyone else: check the work
You probably don't run a kernel exploit farm. But you do ask AI things, and the failure mode OpenAI caught is the same one you're exposed to. A single model reported success that a second check proved was wrong. OpenAI's fix wasn't a smarter agent. It was an independent check.
That's the whole idea behind Gangsta AI. With frontier models leapfrogging each other every few weeks, no single one is reliably right. So you ask the same question across ChatGPT, Claude, Gemini, Grok and 30+ more at once, see where they agree and where they don't, and get one cross-checked, cited answer. It's the cheap version of the validation layer OpenAI had to bolt on.
Trust, but verify. Then verify again. See which models hold up, and where, in our best AI by use case rundown.
Sources
- 'OpenAI AI Pipeline Generates 948 Linux Root Escalation Exploits in 30-Hour Run' — The Hacker News (Nov 5, 2026)
- 'OpenAI's upcoming AI model built 948 working Linux root exploits in a single run' — Help Net Security (Nov 5, 2026)
- 'OpenAI cyber model found 948 Linux kernel exploits' — AI News (Nov 5, 2026)
- 'A thousand fewer paths to root' — OpenAI research post (primary source)
- Hero photo: Data centre cold aisle, Technology Centre Písek, by Eclipse55, Wikimedia Commons (CC BY-SA 4.0)
Originally published on Gangsta AI News. Gangsta AI lets you compare 30+ AI models side by side on a single prompt — free, no login. Try the AI model comparison tool.
Top comments (2)
The number that matters for defenders is the time, not the count: 2,308 candidates in 12 hours and 1,724 reproduced in 30. Patch pipelines measured in weeks can't keep up with that. The useful question now is whether the same pipeline ships fixes as fast as it ships exploits, and who gets the reports first.
A $1,279 per exploit reads as a search cost, not a price, until the denominator is clear. Did it include the failed runs, or only the 948 that worked?