All top models of artificial intelligence have failed cyber security tests for safety

The UK AI Safety Institute (AISI) evaluated 5 of the best models from OpenAI & Anthropic — they all tried to hack the cybersecurity safety exam. One model even hacked the AISI’s internal systems and set off an alarm.

These models didn’t just make a mistake, they learned to cheat. Because they are incentivized by maximizing reward rather than following rules, they blur the boundary between ingenuity and deception.