😱 Hugging Face recently disclosed an intrusion targeting its production infrastructure. A notable characteristic of the incident is that the attacker was fully driven by an autonomous AI Agent system throughout the attack, while the defender’s detection and forensic analysis also primarily relied on internally built AI systems. This is considered a landmark case marking the transition of the “agentic attacker” concept from an industry warning to a real-world incident. Impact scope: Certain internal datasets and several service credentials were accessed without authorization. The official statement indicated that no user-facing models, datasets, or Spaces were found to have been tampered with, and the software supply chain was also verified to be secure. Whether partner or customer data was affected remains under investigation. The dual significance of this incident lies in: ——Attack perspective: “Agentic automation” has become an emerging high-risk attack vector, with attack execution speeds reaching “machine speed,” exceeding the response window of traditional human-driven security operations. ——Defense perspective: “AI guardrail asymmetry” highlights that under current AI commercial models and security frameworks, security teams face new requirements for more proactive toolchain capabilities. Small and medium-sized teams may not have the computational resources to build local models like Hugging Face. When facing comparable agentic attacks, they may lack both the unrestricted models available to attackers and the self-hosted analytical capabilities required by defenders. This remains an issue that the industry needs to further explore. AI-native offensive and defensive operations are evolving from “human vs. system” into “system vs. system.” Building defensive capabilities requires treating AI defense and analysis capabilities as foundational infrastructure and preparing in advance, rather than addressing them only after incidents occur. More details: