Anthropic has deployed new security protocols to prevent competitors from cloning the capabilities of its Claude AI models. These measures have reportedly decreased the volume of data stolen via distillation attacks by 90% to 99%.
The 99% Drop in Harvested Model Data
Model distillation is a process where a smaller, cheaper model is trained using the outputs of a larger, more sophisticated one—essentially allowing a rival to "shortcut" the expensive research and development phase. According to Reuters, Anthropic has successfully curtailed this practice by restricting access to the internal reasoning processes of Claude AI, resulting in the number of extracted reasoning tokens falling to 10 to 100 times lower than previous levels.
This technical barrier is critical because the "reasoning" phase of a model is where the most valuable intellectual property resides. By shielding these tokens, Anthropic is not just protecting a product, but the very logic that gives Claude AI its competitive edge over other large language models .
Zhipu AI and the GLM-5.3 Controversy
The conflict over distillation has taken on a geopolitical dimension, specifically regarding the competition between the United States and China. As Reuters reported, Anthropic has banned tens of thousands of accounts linked to distillation campaigns and explicitly named the Chinese startup Z.ai, also known as Zhipu, as a bad actor. Anthropic claims that Zhipu attempted to distill capabilities from Claude AI to assist in the development of its own GLM-5.3 model.
This accusation elevates the issue from a corporate dispute to a matter of national security. By framing the theft of AI capabilities as a risk to state safety, Anthropic is aligning its corporate interests with broader U.S. strategic goals to maintain a lead in artificial intelligence over Chinese counterparts.
A 0.1% False Positive Rate for Reasoning Protection
To achieve these results, Anthropic implemented a system designed to block proxy networks that rivals use to probe Claude AI repeatedly without detection. The company claims this safety architecture is roughly as robust as the systems used to prevent Claude AI from assisting in the creation of bioweapons. Crucially, Anthropic asserts that this high level of security comes with a false positive rate of less than 0.1%, meaning legitimate users are rarely caught in the dragnet.
The precision of this system suggests that Anthropic has developed a sophisticated way to differentiate between a human user asking a complex question and an automated script attempting to map the model's logic. This capability represets a secondary victory for the company: the creation of a highly effective AI-driven policing tool.
The Shared Anxiety of OpenAI and Google
Anthropic is not alone in this struggle; the report notes that OpenAI and Google share these concerns regarding distillation. The industry is currently witnessing a shift from a phase of rapid, open experimentation to one of guarded proprietary silos. When a rival can use a "teacher" model to train a "student" model, the financial incentive to spend billions on original training data and compute power vanishes.
This trend echoes the early days of the semiconductor industry, where protecting "chip recipes" became more important than the chips themselves. For the AI giants, the goal is no longer just to build the smartest model, but to ensure that no one else can replicate that intelligence for a fraction of the cost.
The Unverified Claims Regarding Zhipu's Methods
Despite the specific accusations, several critical pieces of information remain missing. the source does not provide a response from Zhipu AI regarding the allegations of distilling Claude AI for the GLM-5.3 model, leaving the claim one-sided. Furthermore, it remains unclear exactly how Anthropic identifies "distillation accounts" among tens of thousands of bans, or whether these safeguards might inadvertently hinder academic researchers who use proxy networks for legitimate safety testing.
Comments 0