Papers & Research

Anthropic strengthens defenses against distillation attacks, cutting reasoning token extraction 10 to 100 times

AI summary

Anthropic published a blog post detailing new defenses against Claude distillation attacks, including limiting the flow of reasoning content, using a two-stage classifier to block extraction attempts, and imposing stricter limits on accounts that repeatedly try. Early evidence shows these methods reduced the reasoning data obtained by distillers by about 90% to 99%, and cut the number of reasoning tokens extracted on the platform side by 10 to 100 times. Anthropic also said the measures involve trade-offs, including that some developers need to update integrations and that the classifier's false-positive rate is below 0.1%.

Key Takeaways

  • Anthropic published a blog post detailing new defenses against Claude distillation…
  • Early evidence shows these methods reduced the reasoning data obtained by distillers by about 90%…
  • Anthropic also said the measures involve trade-offs,including that some developers need to update…

Quick Facts

Published
Updated
Source
Anthropic:Research(发表成果 · 网页)
Category
Papers & Research
AI score
77
Read the original Read the full article at the original source
Why it matters

Anthropic 公开其反蒸馏防御的具体做法和效果数据,读者可以据此了解这类防御的成本、权衡和对整个行业的借鉴意义。

Related entities

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: Anthropic strengthens defenses against distillation attacks, cutting reasoning token extraction 10 to 100 times. October 12, 2026. https://new.nat6.net/en/news/lwchujhujmc1tdor61dn39irl/

View all →