Anthropic says new safeguards cut Claude reasoning theft by 90–99%

What Anthropic changed

Anthropic now limits access to Claude’s internal reasoning steps and uses a classifier to detect attempts to extract data. Accounts flagged by the system are blocked automatically.

The company says Zhipu extracted capabilities from leading US models in cybersecurity for its GLM-5.3 model. It also alleges that Moonshot secretly routed requests through Claude. Anthropic warns that stolen capabilities could end up in foreign military and surveillance systems.

A large claimed reduction, with a cost

Anthropic says the measures reduced theft by 90–99% and that it has blocked tens of thousands of accounts. It is urging other AI developers to adopt similar protections, while acknowledging that legitimate users are also being blocked.

I think the percentage is less informative than what it leaves unreported: Anthropic gives no baseline for how much reasoning data was being extracted, or how it measured the reduction. The scale of the bans suggests the controls are being applied broadly, but the company’s own admission about false positives points to a trade-off that other developers would inherit if they copy the approach.

The unresolved tension is whether providers can protect model capabilities without making access less reliable for ordinary users.