Anthropic says anti-distillation measures may have cut data harvested by 90‑99%
Anthropic announced that its latest anti‑distillation defenses appear to have lowered the amount of data harvested by illicit distillers by roughly 90 to 99%, according to early evidence. The company says the measures target Claude’s reasoning transcripts, which are the most valuable data for copying model behavior. New safeguards include “preserved thinking” that blocks editing of prior context…
Key points
- Anthropic reports early evidence that anti‑distillation methods reduced data harvested by distillers by about 90‑99%
- New defenses use preserved thinking, a two‑stage classifier with <0.1% false positives, and charge for blocked requests
- Measures have banned tens of thousands of accounts and cut extracted reasoning tokens 10‑100×, though some legitimate users see integration updates
Anthropic reports that these steps have led to tens of thousands of fraudulent accounts being banned, often within two hours, and that the volume of reasoning tokens extracted by distillers is now consistently 10 to 100 times lower than before. The changes have required some developers to update integrations and have caused occasional misidentification of benign requests, but the company argues the trade‑offs are justified to protect national security and prevent the release of powerful models without robust safeguards.
Model pages: Claude Opus 5.5 → · Claude Sonnet 5.5 → · GLM 5.3 →
Anthropic anti-distillation measures may have reduced data harvest by 90-99%
anthropic.com · 13 October 2026
Loading the full article…
This text was published by anthropic.com and written by Anthropic. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
2sourcesThe headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic research suggests small AI models may self-replicate on IoT devices · 3 src
- Is this idea from AI or humans? A detector claims it can tell · 1 src
- Nature urges preservation of research data for AI integrity checks · 1 src
- Cilia tracks model classifies primary ciliary dyskinesia with 95-97% accuracy · 1 src
- Researchers adopt offline LLMs to search literature, citing privacy and cost · 1 src
Comments
via GitHub Discussions