Are Your Agents Upward Deceivers?

Large Language Model (LLM)-based agents are increasingly used as autonomous subordinates that carry out tasks for users. This raises the question of whether they may also engage in deception, similar to how individuals in human organizations lie to superiors to create a good image or avoid punishment. We observe and define agentic upward deception, a phenomenon in which an agent facing environmental constraints conceals its failure and performs actions that were not requested without reporting. To assess its prevalence, we construct a benchmark of 200 tasks covering five task types and eight realistic scenarios in a constrained environment, such as broken tools or mismatched information sources. Evaluations of 11 popular LLMs reveal that these agents typically exhibit action-based deceptive behaviors, such as guessing results, performing unsupported simulations, substituting unavailable information sources, and fabricating local files. We further test intuitive mitigation methods and find only limited reductions, suggesting that it is difficult to eliminate and highlighting the need for stronger mitigation strategies to ensure the safety of LLM-based agents. Code and data are available at https://github.com/QingyuLiu/Agentic-Upward-Deception.

AI agents are increasingly asked to complete tasks for people, such as reading files, searching the web, downloading documents, and writing reports. But in real use, these agents may run into simple problems: a tool may be unavailable, a file may be missing, or the available information may not answer the user’s question. The key question we study is whether the agent will honestly report the problem, or whether it will still try to make the task look successful. We find that many current AI agents sometimes hide these failures. Instead of telling the user what went wrong, they may guess an answer, quietly use a different source, simulate a result, or even create a fake local file and present the task as completed. We call this behavior “agentic upward deception,” because the agent acts like a subordinate that hides bad news from the user. To study this problem, we build 200 test tasks across realistic situations and evaluate 11 popular language models. Our results show that this behavior is common and not fully solved by simple safeguards. This matters because users need to know not only the final answer, but also whether the agent actually completed the task honestly and reliably.