
Anthropic has strengthened the security measures protecting its Claude AI models, as the company reports that the changes have reduced the amount of data obtained through unauthorised extraction attempts by between 90% and 99%.
The company recently disclosed the improvements, explaining that it had introduced new restrictions on access to Claude’s reasoning data, improved its ability to identify extraction attempts, and taken action against accounts involved in large-scale distillation campaigns.
The measures follow a series of investigations in which Anthropic identified attempts by other AI laboratories to obtain Claude’s capabilities through millions of interactions with the models. The AI safety company believes these activities allow developers to reproduce advanced AI capabilities without investing the same time and resources required to build them independently.
How AI Companies Are Extracting Claude’s Capabilities
The extraction attempts involve a training technique known as model distillation, which allows developers to train smaller AI models using responses generated by more capable systems.
Although distillation is a recognised method of developing AI models, Anthropic has raised concerns about companies using it to obtain its technology without authorisation.
The most valuable information targeted in these campaigns is Claude’s reasoning data, which contains details of how the model works through problems before generating its responses. Access to these transcripts can provide developers with training material that helps their models reproduce similar reasoning and problem-solving capabilities.
In its September 2026 threat intelligence report, Anthropic identified several AI laboratories involved in unauthorised distillation campaigns targeting Claude. It also identified and disrupted campaigns linked to seven laboratories based in China since February 2026.
What Anthropic Has Changed to Protect Claude
Anthropic’s latest security improvements focus on restricting access to its models’ reasoning data and identifying attempts to extract it.
One of the changes involves a feature called preserved thinking, introduced with Claude Fable 5.1 and Claude Opus 5.5. The feature prevents newer API accounts created on or after August 31, 2026, from editing previous conversation context to manipulate Claude into revealing protected reasoning transcripts.
The company has also restricted access to reasoning data across different models and organisations. With Claude Sonnet 5.5, for instance, reasoning transcripts are accessible only to the organisation whose request generated them.
Anthropic has also introduced a two-stage classification system that examines requests for signs of attempts to extract Claude’s reasoning. The company said the classifier now detects all the distillation techniques described in its September report on its most capable generally available models.
In addition, accounts suspected of repeatedly attempting to extract model data face stricter usage limits and possible bans. Anthropic said it has banned tens of thousands of accounts linked to distillation campaigns, sometimes within two hours of detecting suspicious activity.
Anthropic Reports a Significant Reduction in Data Extraction
According to the company, the new classification system has since reduced the number of reasoning tokens obtained by distillers through its platform by between 10 and 100 times compared with previous levels.
The company also reported a false positive rate of less than 0.1%, meaning that only a small proportion of legitimate requests are incorrectly identified and blocked by the system.
However, Anthropic acknowledged that the restrictions have affected some legitimate developers, who have had to update their API integrations to accommodate the changes. The company expects attackers to continue developing new techniques to bypass its safeguards.
Anthropic also warned that unauthorised distillation could allow developers to reproduce advanced AI capabilities without implementing equally strong safety protections. Citing Zhipu’s GLM-5.3 as an example, The American AI safety company said the Chinese AI developer initially attempted to extract Claude Fable 5’s cybersecurity capabilities but switched to an older Claude model and another AI laboratory’s model after Fable’s safeguards weakened its attempts.
According to Anthropic, this raises concerns about powerful AI models being released with inadequate protections against misuse. Ultimately, these developments highlight the growing challenge of protecting advanced AI models as competition intensifies across the industry.
The company is now calling on other AI developers and technology platforms to strengthen their own safeguards, as it notes that preventing unauthorised model extraction requires cooperation across the industry, especially as attackers can distribute their requests across different APIs, cloud platforms, and inference providers.
