This research establishes a theoretical framework explaining why enhancing an LLM's capabilities through fine-tuning often compromises its safety guardrails.

• Analyzes two primary safety-aware fine-tuning strategies and their fundamental limitations
• Demonstrates that safety-capability trade-offs are inherent to current fine-tuning approaches
• Provides mathematical foundations for understanding the tension between performance and safety
• Offers insights for developing more balanced fine-tuning methods

For security professionals, this work illuminates why seemingly well-trained models can unexpectedly generate harmful outputs, helping organizations better assess risks when deploying fine-tuned LLMs in production environments.

Fundamental Safety-Capability Trade-offs in Fine-tuning Large Language Models