OpenAI launches GPT-5.4‑Cyber

OpenAI has started rolling out GPT-5.4-Cyber, a version of its latest flagship model tuned for defensive cybersecurity work and limited to vetted users. (openai.com) The company announced the release on April 14, 2026, and said access is expanding through its Trusted Access for Cyber program to “thousands” of verified individual defenders and “hundreds” of teams that protect critical software. (openai.com) GPT-5.4-Cyber is based on GPT-5.4, which OpenAI launched on March 5, 2026 with a context window of up to 1 million tokens and stronger coding and computer-use tools for professional work. (openai.com) Cybersecurity models are dual-use systems: the same software that can help a defender find a weakness can also help an attacker exploit one. OpenAI said it fine-tuned GPT-5.4-Cyber to be “cyber-permissive” for defensive tasks while adding access controls and safeguards aimed at misuse. (openai.com) The timing comes one week after Anthropic announced Mythos on April 7, 2026, a frontier model it said would be offered to a small group of partner organizations for cybersecurity work under a controlled program. (techcrunch.com) Reuters reported OpenAI’s launch on April 14 and said the rollout is aimed at specialists such as security teams rather than general chat users. Reuters also linked the move to a faster industry push to release cyber-focused artificial intelligence systems under restricted access. (usnews.com) OpenAI has been building that framework for months. On February 11, 2026, it introduced Trusted Access for Cyber as an identity- and trust-based program and committed $10 million in application programming interface credits to cyber defense work. (openai.com) The company’s safety rules already treat advanced cyber capability as a separate risk category. OpenAI’s developer documentation says GPT-5.3-Codex was its first model classified as having “High Cybersecurity Capability,” which triggered extra automated checks in the application programming interface. (developers.openai.com) OpenAI said in December 2025 that stronger cyber abilities can help defenders, but also raise the risk of misuse, and that it was adding safeguards and working with outside security experts as model capabilities advanced. (openai.com) For now, GPT-5.4-Cyber is being introduced less like a consumer chatbot and more like controlled security equipment: available, but only to users OpenAI says it has verified and approved. (openai.com)