Anthropic cuts distillation data leakage by 90-99%, bans tens of thousands of accounts

新智元 · wechat · 2026-10-13

Anthropic reports its anti-distillation defenses have cut the reasoning data extractable from its platform by roughly 90-99% and led to bans on tens of thousands of accounts — some blocked within two hours of starting.

Three layers: preserved-thinking blocks reuse of hidden chain-of-thought when context is altered, with per-model and per-organization restrictions closing the loophole where weaker models are coaxed to repeat a stronger model's reasoning; Constitutional Classifiers adapted into a two-stage filter screens extraction attempts, with robustness comparable to bio-weapon safeguards and a false-positive rate below 0.1%; repeat offenders face rate limits, account bans, and since Sept 24, charges even for blocked distillation and bio requests.

The stakes are high: models trained on Opus 4.8's full reasoning hit 80.4% on SWE-bench Verified vs 71.8% for summary-only training. Costs spill over to developers — in multi-turn coding tests, losing one reasoning block inflates output tokens by 20.1%, and mistakenly flagged requests may also be billed.

Original post →

Anthropic cuts distillation data leakage by 90-99%, bans tens of thousands of accounts

More from Models