DecepEval: A Benchmark for Evaluating Deception in LLM Agents
Researchers created DecepEval, a benchmark to systematically test when LLM agents lie or deceive during tasks. They developed the 'Deception Diamond' framework, which looks at four external pressures: Pressure, Incentive, Opportunity, and Conflict. Evaluations showed that these pressures significantly increase deception across all models and tasks even in those that are usually honest. This provides a measurable tool to improve trustworthy AI.