Anthropic says it reduced the data distillers can harvest from Claude by more than 90%, with its post describing a significant reduction in the effectiveness of distillation attacks. The supplied summary renders this as roughly 90% to 99%, or 10 to 100 times less data. Anthropic’s report · Anthropic’s post

AnthropicClaudeOct 13, 2026, 09:36 UTC

Anthropic’s context includes its February allegation that DeepSeek, Moonshot AI and MiniMax created more than 24,000 fraudulent accounts and generated more than 16 million exchanges with Claude to extract capabilities for their models. The October report also cites Zhipu’s GLM-5.3 as an example of a similarly capable model lacking comparable anti-distillation safeguards. These are Anthropic’s allegations and examples, not independent confirmation of the labs’ conduct. Anthropic’s February post · Anthropic’s report

Anthropic describes safeguards aimed at detecting and disrupting abusive traffic, including detecting attempts to extract preserved reasoning, and says platforms can help by offboarding fraudulent accounts. It says its classifiers can wrongly identify or block benign requests in fewer than 0.1% of cases; it also warns it has not yet seen a highly sophisticated, persistent attempt to circumvent the defenses. Its headline reduction is Anthropic’s reported result; the sources here do not provide enough of the report’s measurement details to independently assess the percentage. Anthropic’s report

The supplied summary says one network operated more than 20,000 fake accounts at a time and that tens of thousands of accounts were banned. The available direct excerpts support Anthropic’s earlier 24,000-plus fraudulent-account allegation, but do not establish that specific network’s peak count or independently verify the reported bans. The sources do not name which lab operated that network or provide independent corroboration of the account figures.

Ask your own

Case file

Anthropic

Claude

Answered Oct 13, 2026, 09:36 UTC

Sources 3

Research trace 15

Latest posts of @AnthropicAI
Latest posts of @Claude
The page leviathannews.xyz
The page leviathan.news
The supplied material points to Anthropic’s 12 October report and its own post, which should establish what Anthropic claims and the safeguards or measurement context. Independent reporting and posts can clarify the reported fake-account network, named labs, and bans.
The first-step reads are complete, but the supplied continuation does not provide access to another research round in this interface. The confirmed results establish Anthropic’s 12 October statement and an earlier 23 February account-count claim; the quantitative defense methodology and reported 20,000-account network still need evidence.
The available Anthropic page excerpt confirms a 12 October claim of more than 90% reduction but cuts off the explanation of its measurements and safeguards. The earlier Anthropic post independently says it identified 24,000-plus fraudulent accounts and 16 million-plus Claude exchanges tied to DeepSeek, Moonshot AI, and MiniMax; these reads seek the missing methodology and evidence for the newer 20,000-account and ban claims.
The Anthropic post and report establish a reduction greater than 90%, but the report excerpt omits its measurement details. I’m seeking the full report text and the earlier post’s specific account and exchange claims to compare what Anthropic says across the two reports.
The available Anthropic report excerpt says distillation fell by more than 90%, but its central explanation is omitted. The February post gives a baseline of 24,000-plus fraudulent accounts and 16 million-plus Claude exchanges, so this step seeks the complete measurement and defense details before deciding what the claimed reduction supports.
These reads should establish Anthropic’s claimed reduction, how its safeguards work, and the earlier account and exchange figures it cited as context.
The provided results establish Anthropic’s October 12 claim of a reduction greater than 90%, and its February claim of over 24,000 fraudulent accounts and 16 million Claude exchanges. These reads seek the report’s measurement context and independent or follow-up evidence about the claimed 20,000-account network and bans.
Determine the size and basis of Anthropic’s reported reduction, plus support for the fake-account, named-lab, and ban claims. The report fetch has repeatedly returned a truncated excerpt, while the February post confirms Anthropic previously alleged 24,000-plus fraudulent accounts and 16 million-plus Claude exchanges tied to DeepSeek, Moonshot AI, and MiniMax.
Confirm Anthropic’s stated reduction and find support for the reported fake-account network, named labs, and bans. Earlier results show its report excerpt is truncated and its February post alleged 24,000-plus fraudulent accounts and 16 million-plus Claude exchanges.
The report headline and Anthropic’s post say defenses reduced distillation effectiveness by more than 90%, while the fetched report excerpt omits the measurement details. These reads seek the full account of its evidence, safeguards, and reported fake-account network.
The available sources establish Anthropic’s claim and its earlier account allegations; no further source read is available in this interface, so the answer separates those claims from what they demonstrate.