OpenAI's GPT‑5.4‑Cyber
OpenAI quietly rolled out GPT‑5.4‑Cyber on April 14 as a model tuned specifically for defensive cybersecurity tasks and is not being added to public ChatGPT today (reuters.com). Access is being gated through a Trusted Access for Cyber (TAC) program that gives initial use to vetted security vendors, organizations and researchers (axios.com). OpenAI also pointed to big gains on capture‑the‑flag style benchmarks across recent releases — performance rising from about 27% on GPT‑5 in August 2025 to roughly 76% on GPT‑5.1‑Codex‑Max, according to early signals (siliconangle.com).
OpenAI has started limited access to GPT‑5.4‑Cyber, a version of its latest model tuned for defensive cybersecurity work, not general public chat. (openai.com) OpenAI announced the release on April 14, 2026, and Reuters reported the launch came one week after Anthropic introduced its own cyber-focused model, Mythos. Reuters said GPT‑5.4‑Cyber is a fine-tuned variant of GPT‑5.4 built specifically for defensive use. (reuters.com) Cybersecurity models are trained to help with jobs like finding software flaws, analyzing malicious code, and suggesting fixes before attackers can exploit them. OpenAI said this release is “cyber-permissive,” meaning it is adjusted to be more helpful for vetted defenders working on real security problems. (openai.com) Access is being routed through Trusted Access for Cyber, an identity-based gate OpenAI first introduced in February 2026 for users handling higher-risk cyber tasks. OpenAI said the expanded program now covers thousands of verified individual defenders and hundreds of teams responsible for protecting critical software. (openai.com 1) (openai.com 2) Axios reported that OpenAI is using a tiered access plan instead of putting the model straight into ChatGPT. Initial users include vetted security vendors, organizations, and researchers, with tighter controls for capabilities that could be misused. (axios.com) OpenAI tied the rollout to a broader claim that frontier models are getting much better at capture-the-flag tests, which are contests where systems solve security puzzles that mimic real attack and defense work. SiliconANGLE reported OpenAI’s early signals showed scores rising from about 27% for GPT‑5 in August 2025 to roughly 76% for GPT‑5.1‑Codex‑Max. (siliconangle.com) The company has been preparing for this inside its main model line too. OpenAI’s March 5 system card for GPT‑5.4 said GPT‑5.4 Thinking was the first general-purpose model in the GPT‑5 series with mitigations for “High” cybersecurity capability. (openai.com) That puts GPT‑5.4‑Cyber in a narrower lane than the standard GPT‑5.4 release from March 5, which OpenAI shipped in ChatGPT, the application programming interface, and Codex as a general professional model with coding and computer-use features. GPT‑5.4‑Cyber is being held back for screened defenders instead. (openai.com 1) (openai.com 2) OpenAI said the program is meant to “help cyber defenders protect us all” as more capable models arrive over the next few months. For now, the company is opening the door a little wider, but only for users it says it can verify. (openai.com)
