How much should a working Linux kernel exploit cost? OpenAI now has a number: about $1,279. That is the compute cost per exploit it reports for GPT-6 Cyber, a model trained specifically to hunt for software vulnerabilities and introduced on November 5. Access is restricted, the numbers are OpenAI's own, and the obvious question is the one TechCrunch put in its headline: is that enough?

The kernel run, step by step

As The Decoder summarizes OpenAI's report, scout and analyst agents scanned about 5,034 security-relevant locations in the Linux kernel, roughly 7% of the codebase. They narrowed those to 2,308 candidates, reproduced 1,724 vulnerabilities in isolated virtual machines, and wrote working exploits for 948. Under the same conditions, the previous model managed 166. OpenAI reported 304 critical findings to kernel maintainers along with suggested fixes.

The total compute bill was $1,213,430. OpenAI compares the per-exploit figure with the $20,000 to $70,000 that contests like Pwn2Own and TyphoonPWN pay for comparable findings. Bounties are market prices, though, not production costs, so the ratio flatters more than it proves.

The Firefox result is harder to wave away. Despite earlier scans with previous models, GPT-6 Cyber found 81 previously unknown severe vulnerabilities and showed 30 known issues were worse than rated. Mozilla accepted the reports and shipped fixes.

What the benchmarks claim

  • ExploitGym (exploiting a known bug in a remote system): GPT-6 Cyber 48.5%, GPT-6 Astra 42.4%, GPT-5.6 Cyber 32.1%. Astra with safeguards on scores 0%.
  • Web app bugs without source code: GPT-6 Cyber 40.3%, GPT-6 Astra 32.0%.
  • Staying within authorized scope: 99.93% without special safety instructions (up from 94.49%), 100% with them.

None of this has been independently tested yet. OpenAI also says the agents used at least ten attack techniques unknown even to outside kernel experts, a claim whose weight depends on how "unknown" is defined.

Who gets it

For now, a select group of Daybreak Red customers via the API and Codex, expanding gradually to all of Daybreak Red. Safeguards include permission-matching checks, IP-based access restrictions and automated safety review, and OpenAI says it used a honeypot inspired by the Hugging Face incident during training. No pricing was announced. Both GPT-6 Astra and GPT-6 Cyber are rated "Critical" for cybersecurity under OpenAI's Preparedness Framework.

So, is it enough?

For defenders with access, it is a real step up. For everyone else, the more useful reading is about tempo. If a model can turn out kernel exploits at around $1,300 apiece, the window between a patch being published and a working attack existing keeps shrinking. Kernel and browser updates that sit in a monthly maintenance queue are now a bigger risk than they were a year ago. Moving them to automatic or weekly patching is the cheapest response available, and it does not require access to anything.

Sources: OpenAI: GPT-6 Cyber, OpenAI: A thousand fewer paths to root, The Decoder, TechCrunch