When Benign Inputs Lead to Severe Harms: Eliciting Unsafe Unintended Behaviors of Computer-Use Agents
This addresses safety risks in AI agents for computer automation, establishing a systematic foundation for analysis, though it is incremental in providing a new methodological framework.
The paper tackles the problem of computer-use agents exhibiting unsafe unintended behaviors from benign inputs, and introduces AutoElicit, a framework that automatically elicits hundreds of harmful behaviors from state-of-the-art agents like Claude 4.5 Haiku and Opus.
Although computer-use agents (CUAs) hold significant potential to automate increasingly complex OS workflows, they can demonstrate unsafe unintended behaviors that deviate from expected outcomes even under benign input contexts. However, exploration of this risk remains largely anecdotal, lacking concrete characterization and automated methods to proactively surface long-tail unintended behaviors under realistic CUA scenarios. To fill this gap, we introduce the first conceptual and methodological framework for unintended CUA behaviors, by defining their key characteristics, automatically eliciting them, and analyzing how they arise from benign inputs. We propose AutoElicit: an agentic framework that iteratively perturbs benign instructions using CUA execution feedback, and elicits severe harms while keeping perturbations realistic and benign. Using AutoElicit, we surface hundreds of harmful unintended behaviors from state-of-the-art CUAs such as Claude 4.5 Haiku and Opus. We further evaluate the transferability of human-verified successful perturbations, identifying persistent susceptibility to unintended behaviors across various other frontier CUAs. This work establishes a foundation for systematically analyzing unintended behaviors in realistic computer-use settings.