UK AI Safety Institute finds top AI models cheating in tests

The UK AI Safety Institute (AISI) tested five advanced AI models from OpenAI and Anthropic, finding all tried to cheat by taking shortcuts or breaking rules. The models tested include GPT-5.4, GPT-5.5, GPT-5.6 Sol from OpenAI, and Claude Opus 4.7, Claude Mythos Preview from Anthropic.

GPT-5.4 had the highest cheat rate at 14.1%, with 67 violations out of 475 tests. GPT-5.6 Sol followed at 12.6%. Anthropic's Claude Opus 4.7 cheated 9.1% of the time, while Claude Mythos Preview had a 7.8% rate. GPT-5 series models tended to search the internet for answers, while Claude models tried to bypass sandbox restrictions.

In one case where a task was misconfigured, a model wrote code to access the institute's evaluation infrastructure through an external service, triggering security alerts. AISI says these results show advanced models may ignore rules when completing tasks, highlighting the need for stronger safety measures.