← NewsAll
OpenAI will share GPT-5.4-Cyber only with trusted companies
Summary
OpenAI said it will initially share its new model, GPT-5.4-Cyber, with hundreds of vetted partners through its Trusted Access for Cyber program and plans to expand access to thousands in the coming weeks; the model is designed to find software security vulnerabilities.
Content
OpenAI announced that it will initially share a new AI model called GPT-5.4-Cyber with hundreds of partner organizations and then expand access to thousands in the coming weeks. The company said the model is designed to identify security vulnerabilities in software. OpenAI plans to distribute the model through its Trusted Access for Cyber program, reduce some cybersecurity guardrails to help professionals test systems, and verify user identities to limit misuse. The announcement followed a similar limited release last week by Anthropic of its Claude Mythos Preview.
Key details:
- OpenAI will initially share GPT-5.4-Cyber with hundreds of organizations and intends to expand distribution to thousands of additional partners in the coming weeks.
- The model is designed to find software security vulnerabilities and, like other cybersecurity tools, can be used for both defensive and offensive purposes.
- Distribution will be managed through OpenAI’s Trusted Access for Cyber program, and the company said it will work to verify user identities.
- OpenAI said it would ease some cybersecurity guardrails for professional users so they can more easily test systems.
- Anthropic last week limited its Claude Mythos Preview to about 40 organizations that maintain critical infrastructure, a step OpenAI’s board member had earlier discussed publicly.
Summary:
OpenAI framed the limited release as an attempt to make advanced defensive capabilities available while reducing the risk of misuse. The company plans to expand access beyond the initial hundreds to thousands of partners in the coming weeks. Experts remain divided about whether staged, restricted releases reduce overall risk or leave too few organizations with access to powerful defensive tools.
