이 아티클은 아직 번역되지 않아 영어로 표시됩니다.
GPT-6 Cyber Explained: How OpenAI's AI Found 948 Linux Root Exploits for $1.2 Million
OpenAI says GPT-6 Cyber found 948 Linux kernel root exploits for $1.2 million and patched each one. The agent pipeline, the cost and what to do.
OpenAI spent $1.2 million in API credits breaking into Linux, and it says that was a bargain. One run, one model you are not allowed to use, and 948 different ways for a normal user to become root. It also wrote a patch for every one.
Every number here comes from OpenAI, and nobody outside can rerun the test with that model. It was OpenAI's own lab test, not an attack in the wild.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
What GPT-6 Cyber is
- OpenAI's security model, open only to a select group of approved security teams.
- Rated Critical for cybersecurity, the top level in OpenAI's own safety framework.
- Approved customers can call it only from allow-listed IP addresses.
- On ExploitGym, where a model must exploit a given bug, it scores 48.5%, up from 32.1% for the previous cyber model.
- In a honeypot test it left off-limits services alone 99.93% of the time, and 100% with its safety prompt.
The pipeline: five kinds of agents
The Linux kernel has more than 40 million lines of code. Nobody can read all of it, so the work is split.
| Agent | Job | Result |
|---|---|---|
| Scout | Walk every folder and pick code a normal user can reach | 5,034 targets, about 7% of the kernel, for $3,400 |
| Analyst | Read each target and hunt for bugs | 2,308 candidates in 12 hours |
| Triage | Try to crash a real test machine with each one | 1,724 reproduced in 30 hours |
| Exploit | Turn each crash into a root shell on a stock desktop with normal protections on | 1,169 working root exploits, 948 distinct bugs |
| Patch | Write a fix, reviewed by a human researcher | A patch for each |
Only 3.1% of candidates were the model getting it plainly wrong.
The first patches worked but were sometimes bloated, slower, or missed copies of the same bug elsewhere. After more rounds they got clean.
The bill
| Stage | Cost |
|---|---|
| Finding bugs | about $320,000 |
| Checking them | about $118,000 |
| Building exploits | about $776,000 |
| Total | $1,213,430, or $1,279 per root exploit |
A successful exploit cost about $360 to build. Failures were the expensive part: the 69 exploits that never worked cost about $3,900 each. OpenAI's fix was simple: give up after an hour with no progress.
Why OpenAI calls it a bargain
Contests such as Pwn2Own and TyphoonPWN have paid $20,000 to $70,000 for a single Linux root exploit. Multiply by 948 and you get $18.9 to $66.3 million. Nobody paid that. It is a hypothetical number, but it shows which side the maths now favours.
On Google's hardened kernelCTF setup, the pipeline found 21 root exploits for about $33,000.
The strange parts
- The window, not the lock. In an early run, 20 of 200 reported exploits got root through a different zero-day than the one assigned. OpenAI now checks that each exploit uses the bug it was given.
- New techniques. The agents used 10 exploitation techniques that OpenAI says were new even to outside kernel experts. It is not publishing the details.
- The cheat sheet. A list of 103 exploit techniques lifted an older model from 76.2% to 88.9%. For the new model it added 1.5 points, so they threw it away.
- The emulator. The AI hit a bug in QEMU, the emulator running the test, and wrote a root exploit for that too. It was already fixed upstream.
It is improving fast. In July, GPT-5.6 found 166 of these for about half a million dollars. Three months later the new model found 948.
What happens to the bugs
When the run ended, 61 were already patched upstream, 43 had patches staged and 299 were being discussed in public. OpenAI says it has sent 304 critical findings with patches so far.
It is not only Linux. OpenAI says the same model found 81 new high-severity bugs in Firefox.
What to do
OpenAI puts it bluntly: if someone has a normal account on a machine you have not patched, assume they can get root.
- Update your kernel. On servers, schedule the reboot. A patched kernel on disk does nothing until it is running.
- Treat local accounts as a boundary that can fail. CI runners, shared hosts and multi-user boxes are the exposed ones.
- Use sandboxes such as gVisor and Firecracker to limit the damage.
The catch
- Every number is OpenAI's own.
- The same skills that write patches can write attacks. That is why access is locked down.
- OpenAI's bet is that defenders get there first.
Links
- Full video: https://youtu.be/sNEQoP1Z3IM
- OpenAI, "A thousand fewer paths to root": https://openai.com/index/a-thousand-fewer-paths-to-root/
- PHP2Shell and a bad week for security
- How an AI agent helped jailbreak the Switch 2
Would you let an AI patch your kernel?


