AI models are starting to cross a new line in cybersecurity. UK AISI just tested OpenAI’s GPT-5.5 and found it reached a similar cyber performance level to Anthropic’s Claude Mythos Preview. On expert-level cyber tasks, GPT-5.5 scored a 71.4% average pass rate, ahead of GPT-5.4 and Opus 4.7. It also completed AISI’s 32-step corporate network attack simulation in 2 out of 10 attempts. That made GPT-5.5 only the second model AISI has seen solve the full attack chain end-to-end.

reactive:frontier-ai-cyber-capabilities