OpenAI launches cyber model
OpenAI released GPT‑5.4‑Cyber, a version of its flagship model tailored for defensive cybersecurity work and initially limited to vetted vendors, organisations and researchers. The rollout uses tiered access and identity‑verification controls rather than a straight public release, with reporting noting restricted availability as the product scales. (reuters.com)
OpenAI has launched GPT-5.4-Cyber, a version of its flagship model tuned for defensive cybersecurity and limited at first to vetted users. (openai.com) The company said on April 14 that it is expanding Trusted Access for Cyber to thousands of verified individual defenders and hundreds of teams that protect critical software. Reuters reported the highest verification tier will unlock GPT-5.4-Cyber, which is being released on a limited basis to security vendors, organizations and researchers. (openai.com) (reuters.com) OpenAI said the new model is “cyber-permissive,” meaning it is trained to allow more legitimate security work that ordinary models often refuse. The company said those tasks include vulnerability research and analysis, and outside coverage said the model can support binary reverse engineering, which is the inspection of compiled software when source code is unavailable. (openai.com) (reuters.com) (xda-developers.com) Cybersecurity work often uses the same prompts as offensive hacking, so a model can struggle to tell defense from misuse. OpenAI said that ambiguity is why it is using identity checks, know-your-customer screening and automated monitoring instead of a simple public rollout. (openai.com 1) (openai.com 2) The move comes one week after Anthropic announced Mythos on April 7 under its controlled Project Glasswing program. Reuters reported Anthropic said that model had already found “thousands” of major vulnerabilities in operating systems, web browsers and other software. (reuters.com) OpenAI has been building the access system for months. It introduced Trusted Access for Cyber on February 5, said GPT-5.3-Codex was its most cyber-capable reasoning model at the time, and committed $10 million in application programming interface credits for cyber defense work. (openai.com) The company said it has evaluated model cyber capabilities since 2023, added cyber-specific safeguards in 2025, and is preparing for “increasingly more capable models” over the next few months. Axios reported OpenAI is trying a different release strategy from Anthropic by widening access through tiers while keeping stronger controls on the most sensitive capabilities. (openai.com) (axios.com) For now, GPT-5.4-Cyber is not a general public product. OpenAI is starting with verified defenders, then using the results of that rollout to shape how later cyber-capable models are tested and released. (openai.com) (reuters.com)
