OpenAI’s cyber model

OpenAI has released GPT-5.4-Cyber, a version of its flagship model tuned for defensive cybersecurity and limited 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 rollout came one week after Anthropic announced its own restricted cyber model, Mythos, on April 7. (openai.com; reuters.com) Cybersecurity work here means finding software flaws before criminals do, like a building inspector spotting weak locks before a break-in. OpenAI said GPT-5.4-Cyber is a fine-tuned variant of GPT-5.4 made to be more “cyber-permissive” for those defensive tasks. (openai.com) OpenAI is not putting this model on open public release. The company said access is restricted through verification, safeguards, and monitoring as it prepares for “increasingly more capable models” in the next few months. (openai.com) That caution follows a week of alarms around Anthropic’s Claude Mythos Preview, which Anthropic also limited to selected organizations under Project Glasswing. Reuters reported that banks and regulators were already weighing the cyber risks from more capable models after Mythos was introduced. (cnbc.com; reuters.com) OpenAI framed its move as a defense program, not a general product launch. In its post, the company said it has been building the effort around “democratized access, iterative deployment, and ecosystem resilience,” while widening use only for qualified security professionals. (openai.com) Other outlets described the release as part of a race between large artificial intelligence labs to control who gets access to models that can spot vulnerabilities more effectively than general-purpose chatbots. Bloomberg reported the new model is meant to be more adept at identifying software security weaknesses, while Axios described a tiered access plan with tighter controls on advanced cyber capabilities. (bloomberg.com; axios.com) The immediate test is whether companies can give defenders better tools without making those same systems easier for attackers to misuse. For now, both OpenAI and Anthropic are answering that by keeping their cyber models behind gates. (openai.com; cnbc.com)