GPT-6 Cyber hits Critical, and OpenAI's $1.2M kernel run shows the bottleneck has moved to triage
OpenAI just rated one of its own models Critical. GPT-6 Cyber, announced Nov 5, sits at that level for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. So the distribution looks nothing like a normal launch: a select group of approved Daybreak Red customers through the API and Codex, with wider Daybreak Red access promised soon.
OpenAI's own numbers: 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber. It stayed in scope on a deliberately tempting honeypot eval in 99.93% of runs without its deployment prompt (94.49% for the predecessor), and 100% with it. That honeypot test was informed by the Hugging Face incident, which tells you what it's for. Mozilla classified 81 of the Firefox findings as high severity and began landing fixes. All of these benchmarks are run by OpenAI, and several are new ones it built itself.
The companion post, "A thousand fewer paths to root", is the more interesting read because it prices everything. A five-stage agent pipeline ran on the Linux kernel in mid-October: scouting $3,426, analysis $316,549, validation $73,330, screening $44,194, exploitation $775,929. About $1.2M total, which OpenAI calls more of an upper bound. Out of that: 2,308 candidate findings, 1,724 reproduced, 948 unique privilege-escalation vulnerabilities proven exploitable. The same harness with GPT-5.6 Sol in July found 166. OpenAI says only 3.1% reflected wrong analysis by the model.
A correction to some of the coverage: several summaries say 948 were patched. They weren't. OpenAI developed 948 candidate patches, and when the 10-day run ended, 61 were patched upstream, 43 had staging patches and 299 were being reported and discussed publicly. The rest go out gradually through Patch the Planet partners, at the kernel community's suggested pace. So OpenAI's own numbers show finding racing far ahead of fixing.
OpenAI says it plainly: the challenge "was not discovery" but deciding what reproduces, what boundary it crosses and what maintainers should see first. That's the part for every builder here. Scanning is now a budget line, and the scarce resource is validated evidence plus maintainer attention. If you run anything built on open source, rescanning as models improve just became a standing cost, and so did patching fast. Posts: https://openai.com/index/introducing-gpt-6-cyber/ and https://openai.com/index/a-thousand-fewer-paths-to-root/
← Back to all articles
OpenAI's own numbers: 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber. It stayed in scope on a deliberately tempting honeypot eval in 99.93% of runs without its deployment prompt (94.49% for the predecessor), and 100% with it. That honeypot test was informed by the Hugging Face incident, which tells you what it's for. Mozilla classified 81 of the Firefox findings as high severity and began landing fixes. All of these benchmarks are run by OpenAI, and several are new ones it built itself.
The companion post, "A thousand fewer paths to root", is the more interesting read because it prices everything. A five-stage agent pipeline ran on the Linux kernel in mid-October: scouting $3,426, analysis $316,549, validation $73,330, screening $44,194, exploitation $775,929. About $1.2M total, which OpenAI calls more of an upper bound. Out of that: 2,308 candidate findings, 1,724 reproduced, 948 unique privilege-escalation vulnerabilities proven exploitable. The same harness with GPT-5.6 Sol in July found 166. OpenAI says only 3.1% reflected wrong analysis by the model.
A correction to some of the coverage: several summaries say 948 were patched. They weren't. OpenAI developed 948 candidate patches, and when the 10-day run ended, 61 were patched upstream, 43 had staging patches and 299 were being reported and discussed publicly. The rest go out gradually through Patch the Planet partners, at the kernel community's suggested pace. So OpenAI's own numbers show finding racing far ahead of fixing.
OpenAI says it plainly: the challenge "was not discovery" but deciding what reproduces, what boundary it crosses and what maintainers should see first. That's the part for every builder here. Scanning is now a budget line, and the scarce resource is validated evidence plus maintainer attention. If you run anything built on open source, rescanning as models improve just became a standing cost, and so did patching fast. Posts: https://openai.com/index/introducing-gpt-6-cyber/ and https://openai.com/index/a-thousand-fewer-paths-to-root/
Comments