OpenAI Grants GPT-5.4-Cyber Access to U.S. CAISI and UK AISI for Cyber Evaluations

Details

• OpenAI is providing access to GPT-5.4-Cyber, a fine-tuned variant of GPT-5.4, to the U.S. Center for AI Standards and Innovation (CAISI) and UK AI Security Institute (UK AISI) for cybersecurity evaluations.
• GPT-5.4-Cyber lowers refusal boundaries for legitimate defensive work, enabling binary reverse engineering to analyze compiled software for malware, vulnerabilities, and robustness without source code.
• Launched alongside expansion of Trusted Access for Cyber (TAC) program, scaling to thousands of verified individual defenders and hundreds of teams protecting critical infrastructure.
• TAC, introduced in February 2026, uses tiered verification; highest tiers get GPT-5.4-Cyber access via identity checks at chatgpt.com/cyber or enterprise reps.
• Model supports advanced workflows like vulnerability research, exploit analysis, and agentic security automation; prior tools like Codex Security fixed over 3,000 vulnerabilities.
• Deployment is limited and iterative to vetted users, preparing for more capable models amid rising AI use by attackers.

Impact

OpenAI's GPT-5.4-Cyber rollout, days after Anthropic's Claude Mythos, differentiates by scaling TAC access to thousands of defenders versus Anthropic's select partners, accelerating vulnerability detection in critical software. This widens defender advantages over attackers exploiting AI, with binary reverse engineering enabling faster analysis without source code. By partnering with U.S. and UK institutes, it aligns with government standards efforts, potentially pressuring rivals to expand similar programs and shifting cybersecurity toward proactive, AI-augmented defenses at scale.