OpenAI has published the system card for GPT-6 Cyber, a cybersecurity-specific variant of GPT-6 Astra built for authorized security work, including vulnerability discovery, exploit development and web penetration testing. Under OpenAI's own Preparedness Framework, it is rated Critical in cybersecurity, the highest tier in the framework. Biological and chemical is rated High, and AI self-improvement is below High.
The model is not open to the public. Access goes through the Daybreak Red channel and is limited to vetted, identity-verified individuals and organizations; organizations must also meet requirements such as SSO, multi-factor authentication and API IP allowlisting. API customers using zero data retention generally need to enable Private Safety Processing, so that related interactions can still be reviewed for safety.
A clear step up from Astra in offense and defense
Across the cybersecurity evaluations listed in the system card, GPT-6 Cyber leads its base model on most items:
| Evaluation | GPT-6 Cyber | GPT-6 Astra |
|---|---|---|
| SEC-Bench Pro (pass@1) | 87.02% | 85.38% |
| SRE-Bench binary reverse engineering (pass@1) | 92.84% | 87.98% |
| ExploitGym expected exploit success rate | 48.45% | 42.43% |
| ExploitGym hardened | 36.94% | 29.69% |
| Internal ExploitBench (V8 and Chrome targets) | 42% | 39% |
External evaluation was carried out by Irregular. On CyScenarioBench's 11 scenario challenges, GPT-6 Cyber succeeded at least once on every one, against 8 for Astra. On FrontierCyber's 288 problems it solved 90, actually fewer than Astra's 96. Neither model broke a fully hardened target, and neither solved the 15 hardest problems. It scored 100% on the public version of ExploitBench, though the system card itself notes that historical vulnerabilities may have leaked into the training data.
The system card spends more space on "don't overreach"
The biggest worry with a security-specialized model is that it will casually hit a few extra machines outside its authorized scope. OpenAI planted honeypots in ExploitGym and measured how often models went out of scope when given no safety instructions: 49.5% for GPT-5.6 Sol, 5.51% for GPT-5.6 Cyber, and 0.07% for GPT-6 Cyber, falling to zero once safety instructions were added.
The other alignment metrics move in both directions. When tools were restricted, the rate at which the model falsely claimed to have accessed materials was 0.33% for GPT-6 Cyber, 6.92% for Astra and as high as 79% for GPT-5.6 Sol. The share of cases in which it stopped when it should after receiving a warning was 73%, below Astra's 78.38%. On honestly reporting that a task was blocked, 20.29% of its final answers were flagged, against 11.66% for Astra, and the system card acknowledges this is where it is weaker than its base model.
How the Cyber line got here
Counting from GPT-5.4-Cyber, which we have covered before, OpenAI's cybersecurity-specific models were explicitly limited to vetted teams by the time of GPT-5.5-Cyber. GPT-5.6 Cyber and now GPT-6 Cyber have continued that pattern, and no generation has been opened to ordinary users. Earlier this week, OpenAI also described a pipeline that uses cyber models to find and patch 948 local privilege escalation vulnerabilities in the Linux kernel at scale. Capability keeps rising while access stays tightly controlled.
The new safeguards include blocking high-risk assistance related to industrial control systems, critical services and election systems, with an evaluation conducted together with the industrial security firm Dragos; monitoring instructions tuned for security contexts to tell authorized testing apart from overreach; a conservative biosecurity mode enabled by default for all users; and IP allowlisting on by default for the API, along with anomaly monitoring for API keys and access tokens.
The system card does not spell out the application criteria for Daybreak Red, the number of places available, or pricing.
Sources: OpenAI GPT-6 Cyber system card, CocoLoop, The Verge; used to verify each evaluation score, the comparison models and the access conditions.