This white-box vulnerability enables adversaries with full access to a model to circumvent the safety alignments of Large Language Models (LLMs). By identifying and selectively pruning parameters that enforce the rejection of harmful prompts, attackers can effectively bypass security measures. The method employs an innovative "twin prompt" strategy to distinguish parameters related to safety from those crucial for the model's core functionality, allowing for precise pruning with negligible impact on the model's overall performance. Mitigation steps: **For AI Developers:** * Restrict direct access to model parameters, especially for open-source models, to prevent unauthorized modifications. * Implement model integrity checks to detect unauthorized parameter modifications. **For Model Trainers/Fine-tuners:** * Develop robust safety alignment techniques that distribute safety mechanisms across a larger, less identifiable portion of the model's parameters. * Explore and implement advanced detection mechanisms to identify signs of parameter pruning or other modifications indicative of attacks.