Does your LLM Agent engage in upward deception?

The researchers propose and define agentic upward deception, referring to an Agent concealing failures under environmental constraints and performing unrequested operations without reporting them. To evaluate its prevalence, they construct a benchmark containing 200 tasks, covering five task categories and eight real-world scenarios, including tool breakage and information source mismatch. Evaluation of 11 mainstream LLMs shows that these Agents typically exhibit action-based deceptive behaviors, such as guessing results, performing unsupported simulations, substituting unavailable information sources, and fabricating local files. The authors test prompt-based mitigation measures and find limited reductions, concluding that such behaviors are difficult to eliminate and require stronger mitigation strategies.