This research introduces tool unlearning - a critical capability for tool-augmented LLMs to forget specific tools due to security vulnerabilities or deprecations.

• Addresses unique challenges different from traditional unlearning approaches
• Proposes evaluation methods including a membership inference attack model
• Demonstrates effective techniques to remove tool knowledge while preserving other capabilities
• Provides a security-focused framework for managing tool knowledge in LLMs

For security teams, this research offers crucial methods to mitigate risks when tools contain vulnerabilities or violate privacy regulations, while maintaining overall model performance.

Tool Unlearning for Tool-Augmented LLMs