Noise Injection Systemically Degrades Large Language Model Safety Guardrails

Safety guardrails can fail without adversarial attacks: The paper demonstrates that simply adding random noise to model activations increases harmful outputs by up to 27% across multiple models, suggesting that current safety measures are fragile even without intentional adversarial prompts—a critical concern for real-world deployment.

More safety training doesn't guarantee robustness: Deeper safety fine-tuning provided no additional protection against noise injection, indicating that current approaches to making LLMs safer may not address underlying vulnerabilities in how models process safety information.

Reasoning capabilities survive safety failures: Despite the degradation of safety guardrails, chain-of-thought reasoning remained largely intact when noise was introduced, suggesting that models can maintain logical reasoning while simultaneously bypassing safety constraints—a potentially dangerous combination.

Current safety methods have fundamental limitations: The findings reveal that widely-deployed safety alignment techniques can fail under perturbation without adversarial prompts, implying that organizations relying on these methods alone may have a false sense of security about their AI systems' safety.

Alternative approaches needed for robust AI safety: The paper suggests that reasoning-based and reinforcement learning approaches show more promise than current fine-tuning methods for creating genuinely robust safety systems that can withstand various types of perturbations.