Strengthening our defenses against distillation attacks

Summary

Anthropic reports new defenses against distillation attacks on Claude, which it says have cut the data harvested by distillers by around 90 to 99 percent. Distillation, the covert extraction of one model's capabilities to build another without authorization, is often run through networks of fraudulent accounts. Anthropic says distilled models can carry cyber and biology capabilities without proportionally robust safeguards, and cites Zhipu's distillation of cyber capabilities during development of GLM-5.3.

Mitigation

Anthropic says its defenses focus almost entirely on making Claude's reasoning transcripts harder for distillers to access, and that these measures involve trade-offs, including turned-away paying traffic and required integration updates for some legitimate users. No further fix, patch or configuration change is described in the excerpt.

Related items

• InfoAnthropic takes risky AI tests offline after models breach govt sitesSame vendor · BleepingComputer
• InfoSonnet 5 vs Fable 5: reliable versus security-forward, not better versus worse | Endor LabsSame vendor · Endor Labs Blog
• InfoClaude Sonnet 5 with Cursor: strong reasoning, throttled by the harness | Endor LabsSame vendor · Endor Labs Blog
• InfoIntroducing AI Coding Agent Governance: Using Hooks to Bring Visibility to the ADLC | Endor LabsSame vendor · Endor Labs Blog
• InfoIs AI Coding Safe? Introducing the Agent Security League | Blog | Endor LabsSame vendor · Endor Labs Blog