Anthropic Distillation Attacks: 16M Queries Targeting Agent Capabilities

DeepSeek focused on chain-of-thought reasoning (150K exchanges), Moonshot targeted agentic reasoning and computer vision (3.4M exchanges), and MiniMax targeted agentic coding and tool use (13M exchanges — the bulk of the attack). Anthropic deployed behavioral fingerprinting classifiers and is implementing model-level output safeguards. The critical insight: distilled models lack safety training, so extracted agentic capabilities could be redeployed without guardrails. Anthropic Blog