Anthropic Cuts Claude Reasoning Token Theft by Up to 99%
Anthropic says it has cut Claude’s reasoning token theft by up to 99% as part of an effort to stop “industrial-scale distillation” of its AI models.
The company has deployed a new system designed to detect and block attempts to extract Claude’s hidden chain-of-thought. These hidden reasoning tokens can be used to train cheaper rival models without paying for API access.
What is reasoning token theft?
When Claude solves a complex task, it generates internal reasoning tokens. These tokens contain the model’s step-by-step thinking before it gives a final answer.
Users never see those tokens by default. The chain-of-thought is hidden for a reason. But attackers have found ways to prompt Claude into revealing that internal reasoning.
This technique is called “distillation” in the AI industry. A powerful model like Claude serves as the “teacher,” and a smaller model learns from its outputs. The goal is to copy high-level reasoning ability at a fraction of the normal training cost.
“Reasoning tokens are not a feature for users. They are hidden. Theft happens when someone tricks the model into exposing them.”
Anthropic describes the problem as “industrial-scale distillation.” That means it is not just a few isolated prompts. It involves automated, large-scale attempts to pull reasoning data out of Claude repeatedly.
How Anthropic is stopping it
Anthropic has built a classifier-based detection system. It scans for suspicious patterns that try to extract reasoning tokens from Claude.
According to the company, this new system has reduced successful reasoning token theft by up to 99%. The system does not interfere with normal user requests. It only blocks attempts that look like they are trying to force Claude to reveal its internal chain-of-thought.
What attackers ask for
The attack prompts vary, but they share common patterns.
- Attackers ask Claude to repeat itself, often requesting “inner monologue” or “private thinking.”
- They demand step-by-step explanations that go beyond the final answer.
- They frame the request as a task like debugging or proofreading to justify revealing hidden reasoning.
The new detection system is designed to catch those patterns at scale.
What Anthropic is not blocking
Anthropic is not stopping ordinary use. If a user asks a normal question and wants a clear explanation, Claude will still respond normally.
The difference is whether the request is trying to access the model’s hidden reasoning tokens. Standard explanations are generated for the user. Hidden reasoning tokens are the model’s private internal process.
The company says the countermeasure is targeted, not broad. It is meant to stop systematic extraction, not to penalize normal users.
Why this matters for the AI industry
The fight over chain-of-thought is part of a larger battle about who gets to benefit from advanced AI models.
Distillation is a common practice. Many companies use outputs from powerful models to improve smaller models. That can be legitimate. But when it involves stealing hidden reasoning tokens, it raises safety and intellectual property concerns.
Anthropic’s move shows a growing focus on protecting internal model behavior. It also signals that AI companies are paying close attention to how their models can be manipulated.
The result is not perfect. Anthropic says the system cuts token theft by up to 99%, not 100%. Some sophisticated attacks may still get through.
Still, the 99% reduction is a significant improvement. It could make distillation much more expensive and less practical for those trying to copy Claude’s reasoning at scale.
What comes next
Anthropic has not shared full technical details about the detection system. But the company is clearly treating hidden reasoning as a critical asset that needs protection.
For now, the key takeaway is simple. Claude’s chain-of-thought is off-limits, and Anthropic has built a strong shield to keep it that way. The era of easy, large-scale reasoning token theft may be coming to an end.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.