Get Started
Home
Topics
Search
Library
Research questionHow can safety evaluations measure harmful actions by computer-using agents rather than chatbot refusals?Chatbot safety tests often assess refusal behavior or response content, but an agent can translate a harmful request into a sequence of computer operations. It is therefore difficult to determine whether the system can actually complete a damaging task.
AI
AI Agents
Alignment & Safety
Evaluation & Benchmarks
Latest papersRecent research connected to this question, newest first.Measuring Harmfulness of Computer-Using AgentsThe evidence comes from CUAHarm, a benchmark of 104 expert-written misuse risks evaluated in a sandbox with rule-based verification of task outcomes. It covers frontier language models and the UI-TARS-1.5 agentic framework; the source also reports results for monitoring unsafe agent actions, including chain-of-thought monitoring and hierarchical summarization.research paper · Sep 3, 2026
Related questions
How should LLM safety be evaluated when harmful prompts vary in implicitness and sophistication?How can safety-tuned language models distinguish harmful requests from benign ones with risky wording?How can safety evaluations measure language-model behavior without triggering evaluation-aware changes in decisions?How can mobile GUI agents assess proposed actions before execution to avoid irreversible harm?