Research question
Can risky agent behaviours be found without building real tool environments?
Key findings
- ToolEmu uses an LLM to emulate tools and an automatic safety evaluator - Found failures in 68% of tested cases for the safest agent
Identifying the Risks of LM Agents with an LM-Emulated Sandbox
Introduced LM-emulated sandboxes for scalable agent risk testing, foreshadowing later agent safety benchmarks.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.