OpenAI has unveiled GPT-5.4-Cyber, a modification of its latest flagship model specifically tuned for work in the field of defensive cybersecurity. This comes after competitor Anthropic announced its advanced AI model Mythos .
OpenAI has unveiled GPT-5.4-Cyber, a modification of its latest flagship model specifically tuned for work in the field of defensive cybersecurity. This comes after competitor Anthropic announced its advanced AI model Mythos .
The Mythos model, announced on April 7, is being implemented as part of Project Glasswing, a managed initiative by Anthropic. The project gives select organizations access to the non-release Claude Mythos Preview model for defensive cybersecurity purposes. It has already identified “thousands” of critical vulnerabilities in operating systems, web browsers, and other software.
OpenAI said that initially, access to GPT-5.4-Cyber will be limited: the model will be made available only to verified security service providers, organizations, and researchers due to its more “lenient” architecture, Reuters writes .
“This is a version of GPT-5.4 that lowers the refusal boundary for legitimate cybersecurity activity and opens up new possibilities for advanced defense processes. In particular, it supports reverse engineering of binary files, which allows security experts to analyze already compiled software for signs of malware, vulnerabilities, and overall system stability without needing access to its source code. Because this model is more ‘permissive’, we are starting its limited iterative implementation among trusted security service providers, organizations, and researchers,” the company said in a blog post.
The company also said it is expanding its Trusted Access for Cyber (TAC) program to include thousands of verified individual cybersecurity professionals and hundreds of teams responsible for securing mission-critical software.
OpenAI is adding new levels to the TAC program launched in February. Higher levels of verification will unlock more powerful model capabilities.
Users approved for the highest level of access will have access to GPT-5.4-Cyber, which has fewer restrictions on performing sensitive cybersecurity tasks, such as research and vulnerability analysis.
