Anthropic Claims New Controls Cut Claude AI Distillation Attacks by Up to 99%

Anthropic has published details of new defensive measures it says have sharply reduced the volume of data being harvested from its Claude AI models through so-called distillation attacks — a technique where outputs from a more capable model are used to train or improve a less capable one.

The company published its findings in a post titled “Strengthening our defences against distillation attacks,” dated 12 October 2026. According to Anthropic, early evidence indicates that the amount of data extracted by suspected distillers has fallen by around 90% to 99% since the measures were introduced. Reasoning-token extraction — where attackers specifically target Claude’s step-by-step thinking processes — reportedly dropped to between one-tenth and one-hundredth of previous levels.

Those are Anthropic’s own figures. No independent technical audit or UK regulatory body has verified them.

What Anthropic Says It Found

The publication follows Anthropic’s earlier disclosure that it had identified what it described as industrial-scale extraction campaigns. The company previously said it had linked activity to Chinese AI companies DeepSeek, Moonshot, and MiniMax, alleging those campaigns generated more than 16 million exchanges through around 24,000 fraudulent accounts. Proxy services were reportedly used to obscure the identity of users and obtain access to Claude at scale.

Distillation itself isn’t inherently problematic. It’s a standard technique in AI development, used legitimately by many organisations to compress or adapt models. But Anthropic argues that unauthorised large-scale extraction can breach its terms of service and allow competitors to reproduce capabilities that took considerable resource to develop.

None of the companies named — DeepSeek, Moonshot, or MiniMax — have provided a public response to Anthropic’s allegations in the available material. Their position on the matter remains unknown.

The Measures Introduced

According to Anthropic’s post, the new controls span several layers. Behavioural classifiers have been deployed to detect patterns consistent with automated extraction. Stricter usage limits are now in place, and accounts showing repeated, persistent attempts to pull reasoning data can be banned. Anthropic also says it has introduced restrictions on access to, and editing of, Claude’s prior context — a feature that distillers had reportedly been using to extract structured outputs at scale.

The company says the goal is to reduce illicit extraction without materially degrading the experience for legitimate customers. That’s a fine line to walk. Behavioural detection systems of this kind can produce false positives, and businesses relying on Claude for high-volume legitimate tasks may find themselves subject to additional checks or temporary limits.

Anthropic hasn’t said how it distinguishes between a legitimate enterprise customer running automated workflows and a bad actor systematically harvesting model outputs. That question remains open.

Scale and Context

Sixteen million exchanges. Twenty-four thousand accounts. Those numbers, if accurate, suggest the problem Anthropic is trying to address is not trivial.

The broader issue is one the AI industry is grappling with collectively. As frontier models become more capable — and more expensive to build — the economic incentive to extract their capabilities through distillation rather than develop them independently grows. Anthropic’s Claude, Google’s Gemini, and OpenAI’s GPT models have all been discussed potential extraction risks, though the specifics vary by company and model generation.

Anthropic has not disclosed what proportion of its total API traffic it believes was associated with distillation attempts, nor has it said whether the measures have had any measurable effect on its commercial revenue or customer base.

Independent Scrutiny

The 90% to 99% reduction figure is striking on its face. But it comes entirely from Anthropic. No external researcher, no UK regulator such as the Information Commissioner’s Office or the AI Safety Institute, and no independent technical body has assessed or confirmed the claim. That doesn’t make it wrong — but readers should weigh it as a company’s account of its own performance.

Asked about the measures, an Anthropic spokesperson has not, at the time of writing, provided additional comment beyond the published post. The company’s post does not specify the time period over which the reduction was measured, which makes the figures harder to assess independently.

So the picture that emerges is this: Anthropic believes it has made meaningful progress against a real and large-scale problem. The technical approach — classifiers, rate limits, account bans — is consistent with standard platform security practice. Whether the gains hold as adversaries adapt their methods is a question the post doesn’t answer.

What This Means for Kent Residents

For businesses, universities, NHS trusts, and public-sector organisations in Kent that use Claude through Anthropic’s API or consumer products, the practical effect of these new controls is worth watching. High-volume or automated use of Claude — even for entirely legitimate purposes — could trigger additional account checks or usage restrictions if the activity pattern resembles extraction behaviour. There’s no evidence that any Kent organisation has been affected, and Anthropic says the measures are designed not to disrupt genuine customers; but anyone relying on Claude for large-scale workflows may want to review their usage against Anthropic’s terms of service.