OpenAI releases cyber model
OpenAI launched a cyber‑specific model, GPT‑5.4‑Cyber, and is limiting access to verified cybersecurity defenders rather than broad public release. Multiple reports emphasize that high‑capability security models will be gated behind verification and trust processes rather than open API access. (siliconrepublic.com (nytimes.com).
OpenAI has started rolling out GPT-5.4-Cyber, a cybersecurity model available only to vetted defenders through a gated access program. (openai.com) The company said on April 14 it is expanding Trusted Access for Cyber, or TAC, to “thousands” of verified individual defenders and “hundreds” of teams that protect critical software. Users in the highest verification tier will get access to GPT-5.4-Cyber. (openai.com) (money.usnews.com) GPT-5.4-Cyber is a variant of OpenAI’s flagship GPT-5.4 model that was fine-tuned for defensive cybersecurity work. OpenAI said it trained the model to be more “cyber-permissive,” meaning it is more willing to help with tasks like vulnerability research that standard models often restrict. (openai.com) (money.usnews.com) Cybersecurity work often means probing software for weaknesses before criminals do. OpenAI said it is pairing stronger cyber models with identity checks, automated monitoring, and tiered permissions instead of offering the same access through its broad public application programming interface. (openai.com) (axios.com) The move came one week after Anthropic announced Project Glasswing on April 7, a program that gives selected organizations access to Claude Mythos Preview for defensive cyber work. Anthropic said launch partners include Amazon Web Services, Apple, Cisco, Google, Microsoft, Nvidia, and Palo Alto Networks. (money.usnews.com) (anthropic.com) OpenAI said its own cyber program has been building since 2023 through a Cybersecurity Grant Program and later cyber-specific safeguards in model deployments. The company said it introduced TAC in February 2026 and is now adding higher verification tiers as model capability rises. (openai.com) (ebs.publicnow.com) OpenAI framed the new model as preparation for “more capable models” it expects to release over the next few months. Reuters reported the initial rollout will be limited to vetted security vendors, organizations, and researchers because of the model’s more permissive design. (openai.com) (money.usnews.com) What is changing is not just the model but the rulebook around it. OpenAI and Anthropic are both treating advanced cyber capability as something to verify first and distribute later, with trust checks standing in for a normal public launch. (openai.com) (anthropic.com)
