OpenAI rolls out GPT‑5.4‑Cyber

OpenAI has released GPT‑5.4‑Cyber, a cybersecurity model that the company is giving first to vetted defenders instead of the general public. (reuters.com) The model is a version of OpenAI’s GPT‑5.4 system that was fine-tuned for defensive security work, according to Reuters and OpenAI’s own announcement on Tuesday, April 14. OpenAI said the rollout sits inside a “Trusted Access for Cyber” programme with staged eligibility rather than open signups. (reuters.com, openai.com) OpenAI said Trusted Access for Cyber uses identity checks and trust-based screening to decide who can use more capable cyber features. The company’s application form says the pilot is for enterprises and security practitioners using the tools for defensive work. (axios.com, openai.com) Cybersecurity models can help defenders find software flaws, inspect suspicious code and speed up incident response, but the same skills can also be misused to plan attacks. OpenAI said the programme is meant to expand access for legitimate defenders while reducing the risk that stronger cyber capabilities end up with the wrong users. (openai.com) The timing reflects a wider shift among artificial intelligence labs toward restricted releases for tools with obvious dual use. Reuters reported OpenAI’s launch came a week after Anthropic announced Mythos, another frontier model being deployed through a controlled access programme rather than broad consumer release. (reuters.com) OpenAI has been building this policy framework for months. In February, the company introduced Trusted Access for Cyber and said it would commit $10 million in application programming interface credits to accelerate cyber defense. (openai.com) The company tied the new rollout to a broader safety system it calls the Preparedness Framework, which evaluates frontier models for risks including cybersecurity and sets safeguard requirements before deployment. OpenAI updated that framework last year to put more weight on real-world controls, governance and disclosure. (openai.com, cdn.openai.com) OpenAI said it began evaluating its models for cyber capabilities in 2023 and started adding cyber-specific safeguards to deployments in 2025. In Tuesday’s post, the company said GPT‑5.4‑Cyber follows earlier security-focused products including GPT‑5.3‑Codex and Codex Security. (openai.com, openai.com) For now, the practical change is narrow: more powerful cyber help is becoming available, but only to users OpenAI says it can identify and monitor. The company is asking interested organisations to apply through the pilot instead of treating GPT‑5.4‑Cyber like a standard consumer model launch. (axios.com, openai.com)