Strengthening our defenses against distillation attacks

# Executive summary Anthropic documents industrial-scale illicit distillation campaigns that used networks of fraudulent accounts and proxy infrastructure to extract Claude's hidden reasoning—enabling foreign labs and potentially military/intelligence actors to replicate advanced cyber capabilities—and describes new defenses (preserved thinking, a two-stage classifier, stricter usage limits/bans, and charging for blocked requests) that reduced harvested reasoning tokens by roughly 90–99% while imposing operational trade-offs.