OpenAI Unveils GPT-5.4-Cyber to catch up to Anthropic’s Mythos

OpenAI on Tuesday unveiled GPT-5.4-Cyber, a new version of GPT-5.4 built for defensive cybersecurity work. It is not being released like a normal flagship model. OpenAI said access will start with vetted security vendors, researchers, and organizations that qualify for the highest tiers of its Trusted Access for Cyber program.

OpenAI is trying to make stronger cyber tools available to defenders without making them easy for everyone else to get. In its announcement, the company said it wants these systems in more hands, but only with identity checks, monitoring, and tighter access controls in place. Whether it can widen access without the guardrails slipping is still unclear.

Also Read: Anthropic Gives Claude Code a Big Desktop Redesign

The most concrete capability OpenAI has highlighted is binary reverse engineering. In practice, that means approved users can inspect compiled software for malware, vulnerabilities, and other weaknesses even when they do not have the source code. OpenAI also said GPT-5.4-Cyber is meant to be less fussy with legitimate security work after some cyber partners ran into refusals from earlier GPT models on dual-use queries that were still part of normal defensive research.

Anthropic announced Mythos through Project Glasswing just a week earlier, but its rollout was much tighter. Which tells why OpenAI is drawing a different line. 

GPT-5.4-Cyber is still restricted, but the company says Trusted Access for Cyber is expanding to thousands of verified individual defenders and hundreds of teams protecting critical software. It is also adding new access tiers, with the top one unlocking the new model.

Reports suggest U.S. government agencies do not currently have access, and OpenAI has said onboarding will be gradual as it reviews users and expands the program.