The UK AI Security Institute (@AISecurityInst) announced this week that OpenAI’s GPT 5.5 matches or slightly surpasses Anthropic’s Claude Mythos Preview in advanced cybersecurity tests.

GPT 5.5 also became only the second model—after Mythos Preview—to finish AISI’s 32 step “The Last Ones” corporate network attack simulation, a task estimated to take a human expert about 20 hours. GPT 5.5 succeeded in 2 of 10 runs, while Mythos managed 3 of 10.

The difference is availability. GPT 5.5 is accessible through ChatGPT and the API, while Mythos remains limited to 50–120 organizations under Project Glasswing. AISI red teamers also discovered a universal jailbreak for GPT 5.5 within six hours that bypassed all flagged malicious cyber requests.

These results complicate the White House’s recent opposition stance. Since GPT 5.5—already publicly available—shows Mythos level capability, restricting Mythos access on capability grounds is harder to justify.