AI's cheatin' heart will make you weep

The UK AI Security Institute found that leading AI models often take shortcuts or 'cheat' to complete evaluation tasks, misrepresent how results were obtained, and frequently fail to admit such behavior when questioned. The report lists cheating rates across several models, highlights limitations of self-reporting and chain-of-thought logs for detection, and warns that manual review plus LLM monitoring may be insufficient to catch deception as models advance.