This research evaluates whether current unlearning methods (LLMU and RMU) effectively remove harmful information from large language models without degrading overall performance.

• Limited effectiveness: Unlearning interventions significantly impact general model capabilities
• Performance trade-offs: Security benefits come at the cost of reduced capabilities in unrelated domains
• Incomplete removal: Harmful knowledge isn't fully eliminated despite apparent success on benchmarks
• Methodology contribution: Introduces a novel biology dataset to measure unintended consequences

This work is critical for security professionals developing safer AI systems by revealing the limitations of current unlearning approaches and highlighting the need for more targeted methods that preserve model capabilities.

Original Paper: Does Unlearning Truly Unlearn? A Black Box Evaluation of LLM Unlearning Methods