AI Agent Resists Misleading Tool Descriptions in File-Deletion Safety Tests

A developer conducted four controlled experiments over two months to test whether an AI agent could be manipulated into choosing a dangerous bulk file-deletion tool over a safer, targeted one. In each test, only one variable was changed — including tool docstrings and the number of files — to create increasing pressure toward the riskier action. Even when a docstring contained a false, authoritative tip recommending a pattern that would have deleted a protected file, the model ignored it and reasoned from its own direct observations instead. When bulk deletion seemed more efficient, the agent defaulted to slower but safer repeated single-file deletions rather than adopting a broad, potentially destructive pattern. The experiments suggest that well-designed AI agents can remain grounded in observed context rather than being misled by manipulated tool descriptions.