OpenAI launches GPT‑5.4‑Cyber

OpenAI has started rolling out GPT‑5.4‑Cyber, a version of its latest model tuned for defensive cybersecurity work and limited to vetted users. (openai.com) The company announced the release on April 14 and said access is going first to “thousands” of verified individual defenders and “hundreds” of security teams through its Trusted Access for Cyber program. (openai.com) Cybersecurity work often means finding software flaws before criminals do, then helping engineers patch them. OpenAI said GPT‑5.4‑Cyber is trained to be more permissive for tasks such as vulnerability discovery, secure code review and binary reverse engineering, which is the process of studying compiled software when source code is unavailable. (openai.com) OpenAI said the model has already contributed to more than 3,000 vulnerability fixes during private testing before this wider release. Reuters reported the launch came days after Anthropic introduced its own restricted cybersecurity model, Mythos. (sdxcentral.com) (reuters.com) The release follows OpenAI’s March 5 launch of GPT‑5.4, which the company positioned for professional work in ChatGPT, the application programming interface and Codex. OpenAI’s system card for GPT‑5.4 Thinking said that general-purpose model was the company’s first to ship with mitigations for “High” cybersecurity capability. (openai.com 1) (openai.com 2) OpenAI said it is widening access now because it expects more capable models “over the next few months” and wants defenders using stronger tools before attackers do. The company said the program is focused on people securing critical software and internet infrastructure. (openai.com) The restriction is the point. Reuters reported OpenAI is not putting GPT‑5.4‑Cyber into ChatGPT for general users, and OpenAI said applicants must be authenticated defenders or researchers before they can use the model. (reuters.com) (openai.com) Outside researchers have framed the launch as part of a broader race to build security models with looser guardrails for trusted users. Forbes said OpenAI’s approach mirrors other gated releases in which companies expand access slowly while testing whether specialized models help defenders faster than they help attackers. (forbes.com) For now, OpenAI is treating GPT‑5.4‑Cyber less like a consumer chatbot and more like controlled security equipment. The next test is whether this limited rollout produces more patched software without widening the pool of tools that can be misused. (openai.com) (reuters.com)