OpenAI gates GPT‑5.4‑Cyber

OpenAI has started rolling out GPT‑5.4‑Cyber, a cybersecurity model that only vetted defenders and selected organisations can use for now. (openai.com) The model is a fine-tuned version of GPT‑5.4 built for defensive work such as finding software vulnerabilities, and OpenAI said it is making the system more “cyber-permissive” for approved users in that setting. Reuters reported the launch on Tuesday, April 14, 2026. (openai.com) (reuters.com) OpenAI said it is expanding its Trusted Access for Cyber program to thousands of verified individual defenders and hundreds of teams responsible for protecting critical software. The New York Times reported the company is sharing the technology only with a group of partners rather than opening it broadly. (openai.com) (nytimes.com) Cybersecurity models are designed to help analysts inspect code, trace how malware works, and spot weak points before attackers exploit them. OpenAI said it is preparing for “increasingly more capable models” over the next few months, and Axios reported the company paired this release with a tiered access plan that decides who gets which tools. (openai.com) (axios.com) The access plan has three levels: broad access to baseline safeguards, trusted access for vetted users who need stronger cyber capabilities, and a restricted tier for the most sensitive capabilities. Axios reported the highest-risk systems could be limited to government agencies or a small set of organisations until safeguards improve. (axios.com) (openai.com) The release comes one week after Anthropic announced Mythos, a cyber model under a controlled program called Project Glasswing. OpenAI and Anthropic are both moving toward narrower releases for security-focused systems instead of putting them straight into public chatbots. (reuters.com) (nytimes.com) OpenAI’s cyber program started in February 2026, before this week’s expansion. Bloomberg reported GPT‑5.4‑Cyber is being offered first to some participants already inside that program. (bloomberg.com) (openai.com) Ordinary ChatGPT users will not be able to switch into GPT‑5.4‑Cyber. CNET reported the model is limited to security experts in the trusted-access pipeline, not the general public. (cnet.com) OpenAI said vetted users in the program can get access to model variants, safety controls, evaluations, and workflows tailored to defensive security work. The company is betting it can widen access for people fixing software flaws while keeping tighter controls on tools that could also help attackers. (openai.com)