Anthropic removes hidden anti-distillation code from Claude Code after its discovery

On July 1, 2026, Anthropic updated Claude Code, its AI-based development tool, and removed hidden program code that the company had used for several months to identify developers trying to siphon data from its models through distillation, including those targeting Chinese AI companies.

What the hidden code did

The code fragment was added to Claude Code several months ago without any public notice: users of the tool knew neither that it existed nor what it did. The mechanism was designed to detect unauthorized distillation of models, a practice in which competitors send mass queries to Claude, collect the responses and use them as a training dataset for their own language models.

Knowledge distillation is, in itself, a legitimate machine-learning technique: a more powerful "teacher" model passes behavioral patterns to a more compact "student" model. This method has allowed a number of Chinese developers to build competitive models at significantly lower cost. However, systematically extracting knowledge from a closed commercial API without permission violates the terms of use and is, in effect, theft of intellectual property.

Key facts:

• The hidden code was added several months ago, without notifying users
• Its function was to identify developers using Claude for illegal distillation
• Chinese AI companies were among the presumed targets
• It was removed on July 1, 2026, after being discovered and made public
• Anthropic confirmed that the code existed and that it had been removed

Why Anthropic removed the code after it became public

When news of the hidden mechanism came out, Anthropic found itself in a contradictory position: a company that consistently promotes the principles of safe and transparent AI had used hidden tracking tools inside its own product for developers. In response, Anthropic removed the code in the July 1, 2026 update and publicly acknowledged that it existed, without disclosing details about the scale of its use or the data collected.

This is the aspect that raises the most serious questions: engineers using Claude Code as a work tool did not suspect that they were taking part, as a side effect, in a corporate counterintelligence operation. The line between protecting a business and abusing users' trust turned out to be thin.

What it means

The incident exposes the escalating confrontation in the global AI race: leading Western labs, including Anthropic, OpenAI and Google DeepMind, face competitive pressure from Chinese companies that release powerful models at significantly lower cost. Suspicions of systematic distillation of Western models have circulated in the industry for a long time.

But the Claude Code episode shows that defensive countermeasures, if hidden from the company's own customers, can do as much reputational damage as the threat they are meant to guard against.