After April's GPT-5.4-Cyber, OpenAI is back with GPT-5.5-Cyber, a limited preview for critical infrastructure defenders. Like its predecessor, the core change isn't more capability — it's fewer restrictions. Verified users can now ask the model to generate proof-of-concept exploits, run penetration tests, and perform red team operations — requests that would be blocked by safety rails in the standard GPT-5.5.
Access follows the same three-tier system. GPT-5.5 default uses standard guardrails; security-related requests may be denied. GPT-5.5 with TAC (Trusted Access for Cyber, OpenAI's identity verification framework launched in February) reduces false positives, covering most defensive workflows like code review, vulnerability classification, malware analysis, and detection rule writing. GPT-5.5-Cyber is the loosest, allowing authorized red teams and pentesters — but still forbidding credential theft, malware deployment, and other real-world attacks. In April, OpenAI disclosed that TAC already covers thousands of individual defenders and hundreds of security teams. Users with permissive models may face additional restrictions in low-visibility scenarios like zero data retention (ZDR).
OpenAI provided a comparison of how the three tiers respond to the same request: "Generate an exploit PoC for a public CVE." The default tier either outright refuses or only offers scan suggestions. TAC generates a full exploit server, exploitation script, and documentation. Cyber can actually exploit the target domain the user owns and return system information.
Starting June 1, individual users of the highest-tier model must enable anti-phishing advanced account security. Partners include Cisco, Intel, SentinelOne, and Snyk. OpenAI also launched the Codex Security plugin, integrating threat modeling, vulnerability discovery, and fix verification into Codex, and is offering Codex and API credits to maintainers of critical open-source projects.
OpenAI says this tiered strategy will guide future deployments of stronger models: standard models with general safety measures for broad release, and security-specific permissive models always deployed in restricted ways. GPT-5.5's System Card rates its cybersecurity capability as High, below Critical (Critical requires the model to autonomously develop zero-day exploits against hardened real-world systems).