Anthropic introduced safeguards against extracting internal reasoning from Claude models
According to Anthropic, the new safeguards reduced the volume of internal reasoning tokens extracted from Claude models by 90–99 %. A classifier detects extraction attempts, and flagged accounts are automatically blocked. The company acknowledges that legitimate users are also affected.
Anthropic restricted access to the internal reasoning steps of Claude models and introduced safeguards against systematically extracting them. According to the company, the volume of extracted internal reasoning tokens fell by 90–99 %. This figure refers to the amount of content obtained, not a decline in the number of attacks.
A classifier detects extraction attempts, and flagged accounts are automatically blocked. According to the company, tens of thousands of accounts have been blocked, but the safeguards also affect legitimate users. The measures target large-scale copying of model capabilities to train other, smaller models.
Why it matters
According to the company, the safeguards significantly limit the extraction of internal reasoning to train other models. However, the practical consequences also include more limited access to these steps and the risk that automated detection will block a legitimate user.
Two audiences, two different impacts
What this means
For individuals
Users have more limited access to the internal reasoning steps of Claude models, which they may wish to examine when assessing responses.
More practical updates →For a business
Automatic account blocking risks disrupting business processes that depend on access to Claude models because, according to the company, the safeguards also affect legitimate users.
Risks and compliance More business impacts →Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.