United Statescountry
DevelopingSecurity

Anthropic says new defenses cut data extraction from Claude models by 90-99 percent

Anthropic told Blockchain.News that new measures reduced data taken from its Claude models by distillation attacks by roughly 90 to 99 percent.

Why it matters

The claims point to pressure on model providers to control bulk querying and account fraud, and to scrutiny of foreign labs accused of distilling U.S. model output. No token, price or market figure is involved in the report. Because a single outlet carries the story and the metrics are Anthropic's own, its independent significance is limited so far.

Key facts

The reduction figure comes from Anthropic's internal platform metrics.

Reported by Blockchain.News

Anthropic reported the reduction on October 12, 2026.

Reported by Blockchain.News

Anthropic described distillation attacks as covert campaigns in which unauthorized parties query a frontier AI model at scale to replicate its capabilities without the developer's consent.

Reported by Blockchain.News

Anthropic characterized the attacks as industrial-scale efforts conducted through networks of fraudulent accounts that let overseas laboratories mimic advanced AI capabilities for a fraction of the development time and cost.

Reported by Blockchain.News

Anthropic said in a September 2026 report that the method has been used extensively by foreign AI laboratories to keep pace with U.S. companies.

Reported by Blockchain.News

Anthropic said distilled models may reach near-frontier capability but ship without proportionally robust safeguards, creating risks in biology and cybersecurity.

Reported by Blockchain.News

Anthropic said Zhipu distilled the cyber capabilities of leading U.S. models during development of GLM-5.3, first targeting Claude Fable 5 before switching to an older Claude model and another laboratory's model after encountering safeguards.

Reported by Blockchain.News

What happened

Anthropic reported on October 12, 2026 that defensive measures it put in place cut the volume of data extracted from its Claude models by distillation attacks by roughly 90 to 99 percent, measured against its own internal platform metrics, Blockchain.News reported. The company describes distillation attacks as covert campaigns in which outside parties query a frontier model at scale to reproduce its capabilities elsewhere, and calls them industrial-scale operations run through networks of fraudulent accounts. Anthropic said it had found such methods used extensively by foreign AI laboratories to keep pace with U.S. companies, which it set out in a September 2026 report. As its stated reason for the concern, Anthropic said distilled models can approach frontier capability while shipping without safeguards that match, naming biology and cybersecurity as risk areas. It cited Zhipu's GLM-5.3 as an example, saying Zhipu targeted Claude Fable 5 during that model's development and moved to an older Claude model and to another laboratory's model after hitting safeguards. Only one outlet covered this story, so the figures, the Zhipu account and the CAISSI assessment of GLM-5.3 rest on Anthropic's own reporting and have not been independently confirmed.

How the story developed

  1. Official statement

    Anthropic reports a 90 to 99 percent drop in data extracted from Claude models

    The company attributes the drop to new defensive measures, measured by internal platform metrics.

    Blockchain.News

  2. First report

    First report by Blockchain.News

    Anthropic Reports 90–99% Reduction in Distillation Attacks on Claude Models

    Blockchain.News

  3. Update

    Anthropic says foreign labs used distillation methods to keep pace with U.S. companies

    The claim comes from an Anthropic report issued in September 2026 and cited in the later coverage.

    Blockchain.News

  4. Update

    Anthropic names Zhipu's GLM-5.3 as an example of a distilled model

    Anthropic says GLM-5.3 development targeted Claude Fable 5 first, then an older Claude model and another laboratory's model after safeguards were encountered. The CAISSI assessment of GLM-5.3 is referenced but cut off in the available text.

    Blockchain.News

Affected entities

Original sources

PublisherReportRolePublished
Blockchain.News
Crypto media
Anthropic Reports 90–99% Reduction in Distillation Attacks on Claude ModelsOriginal