Editorial illustration for Anthropic Cuts Claude Reasoning Token Theft by Up to 99%
Anthropic Cuts Claude Token Theft by 99%
Anthropic says it has slashed the theft of Claude's reasoning tokens by as much as 99 percent, a sign of how aggressive the fight over AI model theft has become. The company's reasoning models generate chains of thought before answering, and those internal steps have turned into a target for rivals looking to train competing systems on the cheap. Rather than building their own models from scratch, some companies have been systematically harvesting Claude's output through networks of fake accounts, numbering in the hundreds of thousands, to copy its capabilities at scale.
Anthropic names two Chinese firms, Zhipu and Moonshot, as participants in this kind of extraction. The stakes go beyond lost competitive edge. Anthropic warns that capabilities pulled from US models this way could eventually show up in foreign military or surveillance systems, raising the issue from a business dispute to a national security concern.
In response, Anthropic has locked down access to Claude's internal reasoning steps and deployed a classifier built to catch extraction attempts in real time, automatically banning accounts it flags. The company says tens of thousands of accounts have already been shut down, and it is now pushing other AI developers to adopt similar defenses.
Anthropic is battling large-scale distillation attacks on its Claude models. In these attacks, companies systematically copy an AI model's capabilities to train their own, smaller models. Anthropic says its new defenses have cut the volume of extracted reasoning tokens, the model's internal chain-of-thought steps, by 90 to 99 percent.
Why this matters
For anyone building on top of Claude, this is a reminder that the economics of frontier models are shifting under constant attack, and that cost shows up somewhere in your API bill or usage terms eventually. Anthropic naming Zhipu and Moonshot directly is notable, too. These aren't anonymous bot farms; they're funded companies with their own model lineups, and the accusation amounts to saying a chunk of China's AI progress runs through scraping Claude's internal reasoning at industrial scale. If true, it reframes the "fast follower" narrative in Chinese AI less as independent engineering and more as structured extraction.
For founders and researchers, the 90-99% reduction claim is worth watching skeptically rather than taking at face value. Anthropic has every incentive to publicize a win here, both to deter copycats and to justify tighter rate limits or account verification that affect legitimate users. Expect the fake-account arms race to escalate rather than end, and expect rivals to make similar claims about their own models being targeted, with similarly thin independent verification.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
From Archives Longtime Wikipedian's "Wake-Up Call" Notes Encyclopedia's Decline Since 2022
From Archives LLMs exceed human agreement on some MT annotation tasks, vary by scheme
From Archives Anthropic study finds small AI models can self-replicate on new devices
Research & Benchmarks
From Archives Anthropic Says Claude AI Accidentally Submitted False Murder Tip
Common Questions Answered
What are reasoning tokens and why have they become a target for AI model theft?
Reasoning tokens are the internal chain-of-thought steps that Claude generates before providing an answer to users. These tokens have become valuable targets for competitors because they contain the model's reasoning process, which can be harvested and used to train competing systems more cheaply than building models from scratch.
How does Anthropic's new defense reduce Claude reasoning token theft by up to 99 percent?
Anthropic has implemented new security measures that cut the volume of extracted reasoning tokens by 90 to 99 percent, making it significantly harder for bad actors to systematically copy Claude's capabilities through distillation attacks. The article does not specify the exact technical mechanisms behind these defenses.
What is a large-scale distillation attack in the context of AI models?
A large-scale distillation attack is when companies systematically copy an AI model's capabilities by harvesting its outputs through networks of fake accounts, then use that stolen data to train their own smaller competing models. This approach allows attackers to bypass the expensive process of building models from scratch by leveraging the intellectual property of established models like Claude.
Which companies has Anthropic directly accused of attempting to steal Claude's reasoning tokens?
Anthropic has named Zhipu and Moonshot as companies engaged in scraping Claude's internal reasoning tokens. These are funded companies with their own model lineups, indicating that a portion of China's AI progress may be running through the systematic extraction of Claude's proprietary reasoning data.
How does token theft impact Claude users and API pricing?
The economics of frontier models like Claude are being constantly attacked through theft, and these security costs ultimately show up in users' API bills or usage terms. This means that the expense of defending against distillation attacks and token theft is passed along to legitimate customers using Claude's services.
Further Reading
- Strengthening our defenses against distillation attacks - Anthropic
- Anthropic says new distillation defenses cut data harvested from Claude by 90% to 99% - Reuters
- Anthropic cuts Claude AI model distillation attempts by up to 90% with new safeguards - Firstpost
- Anthropic releases new defensive measures, reducing data theft from model distillation attacks by 90% to 99% - TechFlow
