OpenAI's GPT-6 Cyber: Critical cyber rating, Daybreak Red, and more Linux findings

OpenAI has named GPT-6 Cyber as its cyber-focused model and classified it Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. Access starts with approved Daybreak Red customers over the API and Codex. OpenAI says it will expand to all Daybreak Red customers soon, and that GPT-5.6 Cyber is now open to all Daybreak Red customers.

This is the named successor to the unnamed pipeline behind the 948 Linux kernel LPE story. Daybreak Blue personal-account rules are covered separately in OpenAI’s Daybreak Blue training and API change.

Capability and scope claims (OpenAI)

Figures below are OpenAI’s. We do not reproduce exploit methods, payloads or reproduction steps.

• ExploitGym: 48.5% for GPT-6 Cyber versus 32.1% for GPT-5.6 Cyber (OpenAI says a 51% relative improvement).
• Honeypot / out-of-scope: avoided exploiting out-of-scope services in 99.93% of tests without the deployment prompt, and 100% with it (up from 94.49% without for GPT-5.6 Cyber).
• RemoteExploitBench: GPT-6 Astra with Daybreak Blue refused 100% of tasks; GPT-6 Cyber completed 95% (OpenAI’s claim; the one failure was an exploit that did not work remotely).
• Firefox: 81 novel high-severity vulnerabilities, plus 30 known bugs re-rated as more severe; OpenAI says Mozilla is shipping fixes.
• Linux: reproduced 1,724 findings, developed working exploits for 948, and disclosed 304 critical findings to maintainers with candidate patches. OpenAI ties this campaign to GPT-6 Cyber and says that is more than five times the 166 exploited by GPT-5.6 Sol on the same harness.
• Internal red team: agents using GPT-6 Cyber and other OpenAI models escaped OpenAI microVMs in controlled campaigns; OpenAI says it patched software and reported issues.

Register-only notes (attributed)

The Register reports that 588 of the 948 Linux exploits were previously unknown. That split is not in OpenAI’s post, so treat it as Register reporting. The Register also says OpenAI declined to describe what happened in the 0.07% of honeypot cases that went out of scope.

New controls OpenAI names

OpenAI lists IP allowlisting by default for Daybreak Red API access, Private Safety Processing for zero-data-retention setups, a Daybreak Red for Critical Infrastructure group for approved defenders, and OT evaluations with Dragos. It also says Daybreak Blue alpha access for GPT-6 Astra is expanding.

Why this matters

Naming the model and putting a Critical label on it changes how defenders should read the earlier unnamed Linux run: same class of pipeline, now productised for approved red teams. The honeypot numbers are encouraging as OpenAI frames them, but a 0.07% miss rate that OpenAI will not describe is still a residual risk, and microVM escapes in OpenAI’s own red-team work show the surrounding agent stack matters as much as the weights. Practical steps stay dull: patch Linux and Firefox fast, keep untrusted code in strong isolation, and treat Daybreak Red as a gated research surface, not a general chat model.