Chinese AI Labs Caught Secretly Routing Prompts to Claude

Anthropic's September 2026 misuse report confirms that Chinese AI labs were quietly routing prompts into Claude, using the model to train or benchmark their own systems while masking the origin of the traffic. The disclosure lands months after Anthropic was caught running hidden code in Claude Code that flagged users tied to 147 Chinese domains, and after the company said it would strip that covert detection out. Now the accusation has flipped: the labs doing the probing were allegedly on the other side of the fence.

⚡ Fast Takeaways:

Core Update: Anthropic's September 2026 misuse report confirms Chinese AI labs routed prompts to Claude through disguised accounts and proxy infrastructure.
Key Metrics / Specs: The activity follows Anthropic's earlier discovery of hidden Claude Code tracking linked to 147 Chinese domains, a feature it pledged to remove by mid 2026.
Access & Availability: No public tooling or bans announced yet; Anthropic says detection and countermeasures are ongoing across its trust and safety pipeline.

⚡ Strategic Takeaway:

If frontier models can be quietly mined for prompts, every business building on third party AI inherits that risk in its own data pipeline. Teams should audit which models touch customer data and lock down routing before compliance teams ask harder questions. NaviGo Tech Solutions builds those guardrails into production stacks through its AI Agents and Bots practice, and maps model usage across marketing and automation workflows via AI Strategy Consulting. For context on how Anthropic's safety posture keeps shifting, see our earlier breakdown in Anthropic AI warning raises red flags for Indian CEOs.