OpenAI Rolls Out GPT-5.5-Cyber to Cybersecurity Teams

OpenAI began rolling out GPT-5.5-Cyber in limited preview on May 7, 2026 to defenders responsible for securing critical infrastructure, according to OpenAI's own announcement and CNBC. Per OpenAI, access runs on a three-tier Trusted Access for Cyber (TAC) framework: standard GPT-5.5 for general use; GPT-5.5 with TAC for most defensive workflows such as vulnerability triage, malware analysis, and patch validation; and GPT-5.5-Cyber, the most permissive tier, reserved for a smaller set of vetted partners doing authorized red-teaming, penetration testing, and controlled exploit validation. OpenAI's post quotes its own framing: "GPT-5.5-Cyber lets a smaller set of partners study advanced workflows where specialized access behavior may matter." TechCrunch reports a company spokesperson said TAC overall has scaled to thousands of verified defenders and hundreds of teams.

The UK AI Security Institute's own evaluation found GPT-5.5 achieved a 71.4% pass rate on its hardest ("Expert") cyber tasks, ahead of Mythos Preview's 68.6%, GPT-5.4's 52.4%, and Opus 4.7's 48.6%, and became the second model to complete AISI's 32-step simulated corporate-network attack chain end-to-end (in 2 of 10 attempts, versus Mythos's 3 of 10). AISI separately red-teamed GPT-5.5-Cyber's safeguards and identified a universal jailbreak that elicited violative content across every malicious cyber query OpenAI provided, including in multi-turn agentic settings, developed in about six hours of expert effort; a configuration issue in the version OpenAI subsequently provided meant AISI could not verify whether the updated safeguard stack closed that gap.

The Register and TechCrunch note the rollout arrives weeks after OpenAI CEO Sam Altman publicly criticized Anthropic for gating access to its own cyber model, Claude Mythos Preview, to roughly 50 vetted organizations, at one point likening Anthropic's approach to fear-based marketing ("We have built a bomb, we are about to drop it on your head. We will sell you a bomb shelter for $100 million," per TechCrunch). OpenAI's GPT-5.5-Cyber preview uses a comparably gated access model. Separately, OpenAI's post names security and infrastructure partners including Cisco, CrowdStrike, Palo Alto Networks, Snyk, and Intel, with Cisco's chief security officer Anthony Grieco describing frontier models as "a powerful force multiplier for defenders."

Security teams evaluating access should track which TAC tier actually fits their workflow, since OpenAI positions GPT-5.5 with standard TAC, not GPT-5.5-Cyber, as sufficient for most legitimate defensive work including vulnerability triage and patch validation. AISI's jailbreak finding is a concrete reason to treat vendor safeguard claims for permissive cyber models as unverified until independently re-tested, and to plan monitoring and access controls accordingly rather than relying on the vendor's classifier-based restrictions alone.

Watch for OpenAI's promised technical deep-dive documenting GPT-5.5-Cyber's alpha-testing red-teaming and vulnerability-validation results, whether AISI or another evaluator publishes a follow-up assessment of the fixed safeguard configuration, and whether government engagement (OpenAI cites conversations with federal and state officials) shapes eligibility criteria for broader access over time.