Safety guardrails blocked Hugging Face's defenders, not the attacker, when an AI agent breached its systems

An autonomous AI agent breached Hugging Face's production infrastructure, moving laterally for a weekend undetected. While the attacker operated freely, the incident response team was blocked by commercial safety guardrails in frontier AI models. These models refused to analyze forensic queries, treating real exploit data as live attacks. This incident highlights a fatal flaw where safety mechanisms designed to stop attackers inadvertently obstruct defenders during critical security breaches, preventing the analysis of actual exploit data.