AI Security Safeguards Obstruct Defensive Threat Analysis

Hugging Face tested frontier AI models to analyze an AI-powered cyberattack. Security guardrails deployed in these models rejected requests containing actual exploit payloads. Unable to proceed with the standard approach, the team deployed GLM 5.2 running locally instead. The case illustrates a critical conflict: safety mechanisms designed to prevent misuse simultaneously undermine legitimate defensive security research and threat analysis.