DETECTING AND PREVENTING DISTILLATION ATTACKS

Summary from the TLDR InfoSec newsletter

Anthropic identified industrial-scale distillation campaigns by DeepSeek, Moonshot, and MiniMax that generated over 16 million exchanges across roughly 24,000 fraudulent accounts to extract Claude's agentic reasoning, tool use, and coding capabilities for training their own models. The labs used proxy services operating “hydra cluster” architectures with thousands of coordinated accounts to circumvent regional access restrictions, with MiniMax pivoting within 24 hours to target new model releases and DeepSeek specifically extracting chain-of-thought data and censorship-safe alternatives to politically sensitive queries. Anthropic has deployed behavioral fingerprinting classifiers, strengthened account verification, and is sharing technical indicators with other labs and authorities, warning that illicitly distilled models strip safety guardrails and undermine export controls.