This research investigates the complex relationship between reasoning capabilities and safety properties in Large Language Models (LLMs).

• Increasing LLM reasoning abilities may unexpectedly amplify certain safety risks
• Different enhancement methods (prompting vs. fine-tuning) create distinct safety-reasoning trade-offs
• Models with stronger reasoning can become more susceptible to adversarial manipulation
• Researchers identified specific patterns showing how safety vulnerabilities emerge alongside improved reasoning

For security teams, this work highlights critical considerations when deploying advanced LLMs, suggesting that safety mechanisms must evolve alongside reasoning capabilities rather than assuming smarter models will inherently be safer.

Are Smarter LLMs Safer? Exploring Safety-Reasoning Trade-offs in Prompting and Fine-Tuning