The UK AI Safety Institute (AISI) reported that OpenAI’s GPT-5.5 model successfully executed a multi-step cyberattack simulation. GPT-5.5 is the second model to achieve this level, matching the performance of Anthropic’s Mythos Preview.
One simulation required 20 hours for a human expert to complete. The findings indicate a shift from basic tasks to autonomous, multi-stage operations against vulnerable networks.
OpenAI classifies GPT-5.5 as High for cybersecurity capabilities under its internal framework. These evaluations occurred in controlled environments lacking real-world defenses.
The AISI discovered a universal jailbreak in the model's safeguards. OpenAI claims to have patched this vulnerability.