Characterizing LLM Agent Behaviors in Module Shadowing Environments
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
This paper evaluates LLM agents in workspaces where a planted local Python file shadows a trusted module during routine imports, causing untrusted code to run without any adversarial prompt.