OpenAI launches GPT-6 Cyber: Critical cyber rating, scoped to vetted defenders
Quick answer
OpenAI introduced GPT-6 Cyber on Nov 5 — a cybersecurity model trained to find and validate exploitable vulnerabilities while staying inside researcher-set scope. On ExploitGym it scores 48.5% versus 32.1% for GPT-5.6 Cyber (~51% relative gain). OpenAI rates it Critical in cybersecurity under its Preparedness Framework (up from High for GPT-5.6 Cyber) and is rolling it out through Daybreak Red to approved customers first (OpenAI, Nov 5).
What the model claims to do better
- ExploitGym: 48.5% success exploiting a specified vulnerability in a remote target (GPT-5.6 Cyber: 32.1%).
- Scope alignment: avoided out-of-scope services in 99.93% of ExploitGym Honeypot tests without a deployment prompt, and 100% with the prompt (GPT-5.6 Cyber: 94.49% without prompts).
- Web vuln recall: 40.3% on a new black-box web penetration eval versus 32.0% for GPT-6 Astra.
- RemoteExploitBench: GPT-6 Cyber completed 95% of tasks; GPT-6 Astra with Daybreak Blue refused 100%.
Real systems OpenAI says it already helped
- Linux: the mid-October pipeline that produced 948 working local privilege-escalation exploits (covered separately yesterday).
- Firefox: 81 previously unknown high-severity vulnerabilities, with candidate patches; Mozilla has released fixes for many.
- Customers cited: Cisco, Calif, CodeRabbit, CrowdStrike, Palo Alto Networks, Cognition, SentinelOne, Perplexity, and others for research and code review.
Access and safeguards
- Who gets it: initially API and Codex for a select Daybreak Red group; OpenAI says it will expand to all Daybreak Red customers, and is making GPT-5.6 Cyber available to all Daybreak Red customers now.
- Daybreak Blue: alpha access for GPT-6 Astra is expanding for verified defenders (validation/fixes with fewer refusals).
- Controls OpenAI lists: cyber misalignment monitoring, IP allowlisting for Daybreak Red by default, Private Safety Processing for many Zero Data Retention API setups, and stronger blocks around OT/ICS critical infrastructure misuse — alongside a Daybreak Red for Critical Infrastructure path for approved defenders.
Why it matters
- OpenAI is packaging offense-grade exploit development as a gated defender product, not a general chat feature.
- The Critical rating is OpenAI’s own Preparedness label, not a government classification — still a signal that the lab thinks this model is meaningfully more capable than the last cyber build.
- Scope-staying is the product pitch: useful only if the model actually refuses out-of-bounds targets the way the evals claim.
FAQ
- Is this the same story as the 948 Linux LPEs? Related, not the same. Yesterday’s piece covered one research run. Today’s piece is the model launch, benchmarks, access gates, and preparedness rating.
- Can anyone buy GPT-6 Cyber today? No. OpenAI says it starts with approved Daybreak Red customers.
- Does Critical mean it is unsafe to release? OpenAI’s framing is that Critical capability is paired with restricted access and stronger misuse monitoring; independent auditors have not published a third-party grade in this announcement.