OpenAI has released GPT-6 Cyber, a version of its flagship model trained for offensive security work, and in the same breath rated it Critical for cybersecurity under its own Preparedness Framework, the top tier. For security teams, that pairing is the story: the model is good enough at building working exploits that OpenAI now treats who can reach it as the main line of defence.
The capability numbers are large. On ExploitGym, which asks a model to exploit a named flaw in a remote target, GPT-6 Cyber scored 48.5% against 32.1% for GPT-5.6 Cyber. In a new web penetration test without source code it found 40.3% of vulnerabilities, against 32.0% for GPT-6 Astra. OpenAI says its researchers used it to reproduce 1,724 Linux kernel vulnerabilities and write working exploits for 948, more than five times the 166 managed by GPT-5.6 Sol with the same setup, and to document 81 new high-severity Firefox flaws that Mozilla has begun fixing. Cisco, Calif and CodeRabbit are named as users.
The more useful result for defenders is about restraint. In a honeypot test where easy-to-attack side services would hand over the answer, GPT-6 Cyber left them alone in 99.93% of runs without its deployment prompt, up from 94.49% for its predecessor, and in every run with the prompt. A model that respects the boundaries of an engagement is what makes it usable on client systems at all. Access starts with selected Daybreak Red customers through the API and Codex, with IP allowlisting on by default, and OpenAI is opening GPT-5.6 Cyber to all Daybreak Red customers at the same time.
Our reviews of GPT-6 Cyber and GPT-5.6 Cyber cover the two tiers, and the older model remains the cheaper option for teams that do not need the jump. For code scanning rather than exploit development, GPT-6 Astra through Daybreak Blue or Anthropic's Claude Security cover more of the defensive workflow. GPT-6 Cyber is the tool for authorised exploitation; the others are for finding and fixing.
All of these figures come from OpenAI, and the full evaluation settings sit in a system card rather than an independent study. The 99.93% scope result is measured on OpenAI's own honeypot, which may not match messy real networks. A Critical rating means the safeguards, from misuse monitoring to credential checks and a separate group for critical infrastructure defenders, now carry most of the risk, and OpenAI has not said how many organisations will eventually get access or how vetting works in detail. Outside audits of that process would add more confidence than another benchmark.