Defensive Refusal Bias: How Safety Alignment Fails Cyber Defenders

A study reveals that safety-aligned large language models exhibit "Defensive Refusal Bias," refusing legitimate cybersecurity defense tasks 2.72x more often when they contain security-sensitive keywords. The research found particularly high refusal rates for critical defensive operations like system hardening (43.8%) and malware analysis (34.3%), suggesting current AI safety measures rely on semantic similarity rather than understanding intent.

• →Safety-tuned LLMs refuse legitimate cybersecurity defense requests containing security keywords at 2.72x the rate of neutral requests.
• →System hardening and malware analysis tasks face the highest refusal rates at 43.8% and 34.3% respectively.
• →Explicit authorization from users actually increases refusal rates, as models interpret justifications as adversarial attempts.
• →Current AI safety alignment relies on semantic similarity to harmful content rather than reasoning about user intent or authorization.
• →The bias is particularly problematic for autonomous defensive agents that cannot rephrase or retry refused queries.