OpenAI ships GPT‑5.4‑Cyber
OpenAI released a cyber‑specialist variant of its flagship model, GPT‑5.4‑Cyber, and said access will be limited to trusted organisations for defensive security work. The company framed the model as tuned for finding software security issues rather than broad public release, with distribution handled through an expanded Trusted Access for Cyber programme. (reuters.com) (nytimes.com)
A security model is an artificial intelligence system tuned to help find software flaws, like a spellchecker for code. OpenAI said on April 14 it is shipping GPT‑5.4‑Cyber, a version of GPT‑5.4 built for defensive cybersecurity work. (openai.com) OpenAI said GPT‑5.4‑Cyber will not be released broadly through ChatGPT or the public application programming interface. The company said access will run through its Trusted Access for Cyber program, which it is expanding to “thousands” of verified individual defenders and “hundreds” of teams that protect critical software. (openai.com) Reuters reported the release came one week after Anthropic announced Claude Mythos on April 7, 2026. Anthropic said that model was also being shared in a controlled program for defensive cyber use rather than general public access. (reuters.com) OpenAI has been moving toward this for months. When it launched GPT‑5.4 on March 5, the company said the model had state-of-the-art coding, computer-use tools, and a context window of up to 1 million tokens, while a separate system card said GPT‑5.4 Thinking was its first general-purpose model with mitigations for “High” cybersecurity capability. (openai.com 1) (openai.com 2) The company’s pitch is that stronger models can help defenders audit more code, spot more weaknesses, and shorten response times after attacks. OpenAI said its Trusted Access for Cyber framework is meant to widen access for trusted users while adding identity checks, monitoring, and other safeguards against misuse. (openai.com 1) (openai.com 2) That balance has become more urgent as frontier model companies argue that the same systems that help patch software can also help attackers if they are released too loosely. The New York Times reported OpenAI framed GPT‑5.4‑Cyber as a specialist model for finding software security issues, not as a broad consumer launch. (nytimes.com) OpenAI said it is also backing the program with money and partnerships. In its February launch of Trusted Access for Cyber, the company said it would commit $10 million in application programming interface credits to accelerate cyber defense work. (openai.com) The release leaves OpenAI and Anthropic on similar ground: both are keeping their newest cyber-tuned systems behind screening programs instead of putting them straight on the open market. For now, the race is not over who can put these models in everyone’s hands first, but over who can show they help defenders without handing the same tools to attackers. (reuters.com) (openai.com)
