OpenAI’s new cybersecurity AI scores 85.6% on a key benchmark, surpassing Anthropic’s banned Mythos model amid U.S. security concerns.
OpenAI released GPT-5.5-Cyber, a cybersecurity-focused AI model designed to identify and fix software vulnerabilities. The model achieved an 85.6% score on the CyberGym benchmark, outperforming Anthropic’s Mythos 5, which scored 83.8% before being banned by U.S. authorities.
Anthropic’s Mythos 5 was pulled offline on June 12 following an emergency export control directive citing national security risks. The ban stemmed from concerns over a jailbreak technique that could bypass the model’s safety protocols. OpenAI’s model, part of its Daybreak cyber defense program, remains accessible to trusted defenders under controlled conditions.
The benchmark, developed at UC Berkeley, tests AI agents on 1,507 known vulnerabilities across 188 open-source projects. OpenAI’s release contrasts with Anthropic’s restricted approach, highlighting differing regulatory and security responses in the AI sector.