OpenAI Launches GPT-5.5-Cyber Security Model with Reduced Guardrails for Vetted Defenders
Tags AI · Enterprise · Infrastructure
OpenAI launched GPT-5.5-Cyber, a specialized variant of GPT-5.5 with reduced cybersecurity guardrails, available in limited preview through its Trusted Access for Cyber program to vetted defenders. Unlike standard GPT-5.5 which refuses exploit requests, GPT-5.5-Cyber can generate proof-of-concept exploits, run attack simulations, and validate vulnerabilities by launching simulated attacks against test systems. The model scored 81.9% on the CyberGym benchmark (1,500+ historical vulnerabilities). Launch partners include Cisco, CrowdStrike, Palo Alto Networks, Cloudflare, Intel, Snyk, and SentinelOne. Starting June 1, 2026, individual users on the highest access tier must enable phishing-resistant authentication. The launch is a direct response to Anthropic's Claude Mythos.
Technical significance
GPT-5.5-Cyber represents a fundamental tension in AI safety: the same capability that makes AI valuable for defense makes it dangerous in offensive hands. The tiered access model is an experiment in capability gating that the entire industry will watch. If successful, it could become the template for responsible release of dual-use AI capabilities.