
OpenAI shows how its agents found 948 Linux flaws in a single run
Redacción IvokaNov 5, 7:30 AM 3 min
An agent system using a cybersecurity model not yet released found, tested and generated patches for 948 Linux kernel vulnerabilities. The run cost about $1.2 million in API credits.
OpenAI published a research report describing an automated agent system that found, verified and generated patches for 948 privilege-escalation vulnerabilities in the Linux kernel (the foundation of many servers and systems). It did so with a cybersecurity model that has not yet been released.
How it worked
- A first group of agents picked the security-relevant parts of the code, leaving about 7% to audit.
- The system analyzed 2,308 candidate findings and reproduced 1,724 vulnerabilities.
- After grouping those traced to the same flaw, 948 distinct vulnerabilities remained, each with its own patch.
- In a sample of 166 exploits, the agents used 10 techniques that, to the researchers' knowledge, had not been seen before.
- The run took place in mid-October and cost $1,213,430 in API credits.
What happened with the patches
The patches the system generated are not the same as fixes already applied. When the analysis wrapped up after 10 days, 61 of the privilege-escalation bugs had been patched upstream (in the kernel's main project), 43 had staging patches and 299 were being reported and discussed publicly. OpenAI says it has sent hundreds of reports and patches so far, starting with the most serious flaws, and that security researchers review every exploit, patch and proposed severity rating before anything goes to kernel maintainers.
A warning from the report itself
The results the agents reported did not always hold up: in an early run, the system claimed 200 successful exploits, but human review showed that 20 had reached root through a different flaw than the one assigned to the agent. So, according to OpenAI, even successful agent work needs proper validation. OpenAI also says this kind of testing is relatively inexpensive for defenders.
Why it matters
You don't need to use Linux to notice the signal: agents can already do long, multi-stage jobs like this one, which runs from searching for flaws to verifying them and proposing patches.

