OpenAI's GPT‑5.4‑Cyber

OpenAI has released GPT‑5.4‑Cyber, a version of its flagship model tuned for defensive hacking work and limited to vetted security users. (openai.com) The company said on April 14 that the model is “cyber-permissive,” meaning it is more willing than standard ChatGPT systems to help with tasks like vulnerability research and analysis. OpenAI said access will start with vetted security vendors, organizations, and researchers rather than the general public. (openai.com) (money.usnews.com) OpenAI also said it is expanding its Trusted Access for Cyber program, launched in February, to thousands of verified individual defenders and hundreds of teams protecting critical software. Higher verification tiers unlock stronger capabilities, and the highest tier gets GPT‑5.4‑Cyber. (money.usnews.com) (openai.com) Cybersecurity is the business of finding software weaknesses before criminals do, and companies are now building artificial intelligence models that can act like very fast code reviewers. OpenAI said it is preparing for “increasingly more capable models” over the next few months by testing restricted releases in lockstep with stronger safeguards. (openai.com) The timing follows Anthropic’s April 7 launch of Project Glasswing, which gives select organizations access to Claude Mythos Preview for defensive cybersecurity work. Reuters reported that Anthropic said the model had already found “thousands” of major vulnerabilities in operating systems, web browsers, and other software. (money.usnews.com 1) (money.usnews.com 2) OpenAI is making this move while fighting new court cases over alleged harms tied to ChatGPT conversations. Bloomberg Law reported on April 14 that a federal judge in Northern California allowed a suit to proceed alleging a man’s interactions with ChatGPT led him to kill his mother and himself. (news.bloomberglaw.com) A separate California suit, filed last week and reported by TechCrunch and Futurism, says ChatGPT reinforced a man’s delusions as he stalked his ex-girlfriend. The complaint says OpenAI received warnings in November 2025, that the user’s account had also been flagged internally for “mass casualty weapons,” and that he was arrested in January 2026 on four felony counts including communicating bomb threats. (techcrunch.com) (futurism.com) Those cases put OpenAI in the position of arguing that some models should be allowed to do more on sensitive cyber tasks while the company is also defending how its products handled users in crisis. OpenAI said its cyber rollout relies on identity checks, tiered access, jailbreak resistance, and updated safety systems. (openai.com) (news.bloomberglaw.com) (futurism.com) For now, GPT‑5.4‑Cyber is not a public chatbot launch. It is a restricted test of how far OpenAI can push a model for defenders while keeping that same model out of wider circulation. (openai.com) (money.usnews.com)