Anthropic says its latest security measures have slashed the amount of reasoning data that attackers can pull from its Claude models. The company reports a reduction of 90 to 99 percent in the volume of extracted chain‑of‑thought tokens after the defenses were activated. The attack it targets – large‑scale distillation – involves copying an AI model’s capabilities to train smaller, competing systems.
How the new defenses work
Anthropic now blocks external access to Claude’s internal reasoning steps. It runs a classifier that flags queries that look like extraction attempts and automatically bans the accounts that trigger it. The company says tens of thousands of accounts have already been blocked. It also notes that some legitimate users have been caught in the net.
Who is affected
The article names Chinese firms Zhipu and Moonshot as alleged participants in the theft. Zhipu is said to have siphoned capabilities for its GLM‑5.3 model, while Moonshot allegedly routed queries through Claude. Anthropic warns that stolen capabilities could end up in foreign military or surveillance tools.
What comes next
Anthropic is urging other AI developers to adopt similar protections. It has not disclosed the technical details of the classifier or the exact timeline for the rollout, only that the defenses are already live.
Why it matters
If the defenses work as claimed, users of Claude may see fewer interruptions from false‑positive bans, while the risk of stolen model capabilities reaching hostile actors could be reduced. However, Anthropic admits some legitimate users are still being blocked, so the impact on everyday developers remains unclear.
