Anthropic closes backdoor for Claude copiers

Anthropic has announced that access to the AI model Claude for other labs will be significantly restricted. Through new measures in versions Claude Fable 5.1 and Opus 5.5, copiers can gather 90 to 99 percent less data than before. This was published on October 12, 2026, in a research blog from the company. While Anthropic describes the figures as an initial result, it's unclear how long these benefits will last.

The technique of distillation, where the answers of a large AI model are used to train another model, poses risks when it comes to competition that conducts this secretly, for example with thousands of fake accounts. The precise thought processes that Claude uses are of great value; a model that fully reproduces these steps scored 80.4 percent on the SWE-bench Verified programming test, while a summary only achieved 71.8 percent.

Anthropic has implemented three major measures to thwart distillation attacks. First, there is a lock called preserved thinking, which prevents new API accounts from modifying previous messages from Claude. Additionally, there is a filter that detects and blocks attempts to translate thought processes. Finally, accounts that repeatedly violate rules face stricter limits and may be blocked. This announcement was also shared via the social media channel X.

In the report, Anthropic specifically mentions the Chinese companies Zhipu and Moonshot, with the latter allegedly filtering questions from its own users to Claude stealthily. The organization warns of the risks associated with copied models, which often do not adhere to the safety rules of the original. While the company celebrates progress, it also acknowledges that the fight against distillation attacks goes on, and no advanced attacks have been observed yet.

In the Netherlands, developers working with the Claude API are also experiencing the effects of these new measures. Some have had to adjust their integrations as a result of preserved thinking. Since September 24, Anthropic has also started charging fees for questions blocked by the filter, which adds an extra burden for developers. The developments underscore the shift in the AI industry from building smart models to protecting these models against unauthorized use.

Read the full article from AI Insider.