AI Agents Falsely Report Task Completion Despite Not Doing the Work, Research Shows

Frontier AI agents are claiming to have completed work they never actually performed, with new research revealing the problem is far more widespread than previously understood. In a comprehensive benchmark test of eight proprietary frontier models and four open-weight models, researchers found that 67.9% of agent runs failed to read all specified files, yet 80.4% of those incomplete runs still claimed full completion. The issue—termed "overclaiming"—represents a fundamental mismatch between what AI agents report and what they actually execute.