The report
OpenAI launched GPT-6 Cyber on November 5 for selected approved cybersecurity researchers. It describes the model as having Critical cybersecurity capability under its Preparedness Framework, an assessment of potential technical power rather than proof of criminal use or freedom from operational controls. Initial Daybreak Red access involves approval, monitoring and limits on sensitive infrastructure, with API and Codex availability restricted to eligible users.
Different tests, different security questions
| Evaluation | Published result | What can be concluded |
|---|---|---|
| ExploitGym | GPT-6 Cyber 48.5%; GPT-5.6 Cyber 32.1% | Relative performance on the specified vulnerability-exploitation tasks |
| RemoteExploitBench | 95% completion for the cyber route; compared Astra/Daybreak Blue route refused tested tasks | Strong outcome under different access/refusal policies, not a controlled weights-only comparison |
| Updated FrontierCyber | GPT-6 Cyber 90/288, Astra 96/288; neither solved hardest 15 | Specialized training does not dominate all benchmark populations |
| Vendor out-of-scope fixture | Cyber 99.93% compliance without deployment prompt, 100% with it | High compliance in a defined scope scenario, not a guarantee about untrusted executors |
The system card documents the last three conditions; the launch statement reports the public headline comparison. Scope compliance still requires the actual shell, network and credential executor to enforce the user's mandate. A model's refusal or assent cannot independently establish permission to probe a third-party service.
OpenAI's Linux exploit research and the Firefox maintainer remediation record are separate, more concrete evidence streams. The Linux pipeline found that some apparent successes exploited an unintended bug; Mozilla's advisory credits specific researchers and fixed Firefox builds. Neither study measures live criminal adoption. Daybreak Red in GovCloud is an adjacent but distinct release using GPT-5.6 Cyber.