OpenAI gates cyber model
OpenAI began rolling out a specialized cybersecurity model, GPT‑5.4‑Cyber, but is restricting access to vetted security vendors, organisations and researchers under a tiered Trusted Access programme. The company says onboarding will be gradual and performance on capture‑the‑flag-style tasks has jumped across recent model versions, according to reporting. (axios.com, siliconangle.com)
OpenAI has started rolling out GPT‑5.4‑Cyber, a cybersecurity model that only vetted defenders can use at first. (openai.com) The company said on April 14 that the model is a “cyber-permissive” variant of GPT‑5.4 tuned for defensive work, and that access will begin with vetted security vendors, organizations, and researchers. Reuters reported the first rollout is limited because the model is more permissive on sensitive security tasks. (openai.com, tech.yahoo.com) OpenAI is also widening its Trusted Access for Cyber program, or TAC, from a February pilot to thousands of verified individual defenders and hundreds of teams that protect critical software. The company said higher verification tiers unlock more capable tools, with the top tier getting GPT‑5.4‑Cyber. (openai.com, openai.com, tech.yahoo.com) Cybersecurity models are software assistants trained to inspect code, trace bugs, and test systems the way a human security team would. OpenAI said those same skills can help defenders find and fix flaws faster, but can also be misused if the tools are handed out broadly without checks. (openai.com) OpenAI has been building those checks in stages since 2023 through its Cybersecurity Grant Program, cyber evaluations, and its Preparedness Framework. In February, it added Trusted Access for Cyber and committed $10 million in application programming interface credits to defense work. (openai.com, openai.com, openai.com) The timing follows a similar move by Anthropic. Reuters said Anthropic announced Claude Mythos on April 7 under a controlled program called Project Glasswing for select organizations doing defensive cybersecurity work. (tech.yahoo.com) OpenAI has been warning for months that cyber performance is rising quickly. In a December 2025 post, it said capture-the-flag benchmark scores climbed from 27% on GPT‑5 in August 2025 to 76% on GPT‑5.1‑Codex‑Max in November 2025, a sign that newer systems are getting much better at security challenge tasks. (openai.com) The company said its answer is not a public launch inside ChatGPT, but a slower identity-based program with know-your-customer checks, monitoring for suspicious activity, and tighter access for the most capable models. That leaves GPT‑5.4‑Cyber as both a product release and a test of whether frontier cyber tools can be distributed in narrow lanes before they spread wider. (openai.com, openai.com, axios.com)
