On Tuesday, OpenAI published “A thousand fewer paths to root,” describing what a pipeline of agents found in the Linux kernel. The numbers: 2,308 candidate findings analyzed, 1,724 vulnerabilities dynamically reproduced, 291 proof-of-concepts for information leaks, and 948 working local privilege escalation exploits — the path from an ordinary user account to root.
The run happened in mid-October, used an upcoming cyber model, and cost $1,213,430 in API credits. That works out to $1,279 per exploit.
How the pipeline works
It first identifies the security-relevant parts of the kernel, statically scans those audit targets for candidate findings, then dynamically reproduces the findings in an isolated environment. From there it builds a proof-of-concept to demonstrate exploitability and generates a candidate patch. A human security researcher reviews the output at the end. What comes out is a root cause analysis, a minimal crashing PoC, a patch candidate and a full exploit.
One detail makes the numbers readable: the target VMs ran a stock desktop operating system, not a weakened kernel, under the same threat model used by Pwn2Own and TyphoonPWN. Those competitions have historically paid $20,000 and $70,000 for Linux kernel LPE exploits. Applied to 948 exploits, OpenAI gets a hypothetical valuation of $18.9 million to $66.3 million, roughly 16x to 55x the cost of the run.
On an adapted kernelCTF environment with reduced attack surface, the pipeline produced 25 information leak PoCs and 21 LPE exploits for about $32,936. For comparison, kernelCTF has accepted 26 confirmed 0-day LPEs this year. OpenAI says the agents discovered and used at least ten novel Linux exploitation techniques.
What OpenAI takes from it
The reports and patches aren’t all going out at once. Following guidance from the kernel community, OpenAI started with the critical vulnerabilities and is working with partners in its Patch the Planet initiative on further reports. Back in June, GPT‑5.5 Cyber built 24 LPE exploits under the same program.
Two observations are more interesting than the headline count. First: during the ten days the scan was running, more than 100 Linux LPEs were patched or had staging patches from other contributors, with roughly 300 more under public discussion. The community moves fast on its own. Second: the October run with the newer model found more than five times as many novel LPEs as a GPT‑5.6 Sol run in July, despite three months of hardening in between.
The price per exploit is the actual story
That a model can find kernel bugs has been known since summer. The new number is $1,279. As long as a working exploit took weeks of specialist labor, scarcity was the defense. At four figures it isn’t, and it stops being one for both sides at the same time: the arithmetic that funds OpenAI’s patching program funds anyone with the opposite goal just as well.
OpenAI’s conclusion is that automated patch generation is now essential, because you can’t keep up with the discovery rate any other way, and that large projects need several rounds of scanning and fixing, ideally with each new model generation. As arithmetic, that convinces me. As a forecast, it worries me. A project that can’t afford those rounds is left with the same kernel everyone else can now scan.