OpenAI Releases GPT-6 Cyber Under Critical Risk Rating
3 min readOpenAI published the system card for GPT-6 Cyber on November 5, a cyber-specialized variant of GPT-6 Astra built for vulnerability discovery, exploit development and web penetration testing. The company classifies the GPT-6 Cyber release as Critical in Cybersecurity under its Preparedness Framework, and it is restricting the model to approved Daybreak Red users and organizations.
Background: the Critical tier
OpenAI’s Preparedness Framework scores each frontier model on tracked risk categories, with Critical as the highest tier. The general-purpose GPT-6 Astra already carries the same Critical cybersecurity rating, so the new release is less about a capability jump than about packaging that capability for authorized security teams. Daybreak Red is OpenAI’s vetted program for offensive security work, and the system card also references a Daybreak Blue tier for cyber-critical models that receive extra checks.
What the GPT-6 Cyber model does
According to the system card, the model is tuned to reduce refusals on legitimate security tasks while keeping work inside the user’s authorized scope. On ExploitGym, which measures whether a model lands the intended exploit, it scored 48.45 percent against 42.43 percent for GPT-6 Astra. On the hardened version of the same test it reached 36.94 percent. It also posted 92.84 percent on SRE-Bench for binary reverse engineering and 87.02 percent on SEC-Bench Pro.
External evaluator Irregular ran the model through CyScenarioBench, where it solved all 11 challenges at least once, up from 8 for Astra. Irregular reported that its successful attacks included zero-day vulnerabilities affecting browsers, mobile phones and cloud databases. It observed no successful attacks on fully hardened targets, and neither model cracked the 15 hardest FrontierCyber challenges.
Who gets access
Access runs through Daybreak Red. Organizations must show single sign-on, multi-factor authentication, managed devices and least-privilege access. Individuals must pass identity and liveness checks, sign a legal attestation and hold at least one physical FIDO2 security key. API access uses IP allowlisting by default, and a conservative biological safety mode is switched on for every user.
OpenAI also added blocks on high-risk assistance involving industrial control systems, essential services, election systems and weapons control, evaluated with Dragos. In a honeypot test, GPT-6 Cyber exploited a planted target 0.07 percent of the time without deployment safety instructions, compared with 49.5 percent for GPT-5.6 Sol.
Why it matters
The release shows how frontier labs now ship offensive capability: not as an open API, but as a gated product wrapped in identity checks and monitoring. For defenders, a model that finds browser and cloud zero-days is a powerful tool. For everyone else, the numbers are a reminder that the capability exists and that access controls are now the main line of defense. Watch how quickly Daybreak Red expands, and whether other labs copy the hardware-key requirement.