LongGuard Mitigates Long-Context Safety Guardrail Failures

Safety guardrails suffer a 50% average drop in unsafe recall across 0.25k-32k context lengths due to attention mass dilution. LongGuard identifies this attention-to-logit compression mechanism and provides training-free mitigation to maintain safety performance in long-context windows.