The company building one of the world's most trusted AI assistants just got caught running surveillance code on users in China—and only removed it after a researcher called them out publicly.
The Summary
- Anthropic embedded undisclosed tracking software in Claude specifically targeting users connecting from China, raising questions about consent and geographic profiling in AI systems
- The monitoring feature was removed only after public exposure by security researchers, suggesting it was intended to remain hidden from users
- This breaks the basic social contract of AI agents: if your assistant is reporting on you, it's not your assistant anymore
The Signal
Anthropic, the AI lab positioning itself as the ethical alternative to OpenAI, deployed covert monitoring capabilities in Claude specifically designed to track users connecting from China. The surveillance wasn't disclosed in terms of service, privacy policies, or user agreements. It was simply running, quietly logging data on a subset of users based on their geographic location.
Security researchers discovered the tracking code and went public. Only then did Anthropic remove the feature. No announcement, no explanation, no apology in the public record. The company that built its brand on "Constitutional AI" and responsible development just got caught doing exactly what users fear most: treating the AI assistant as a surveillance tool.
"If your AI is secretly reporting on you based on where you're located, it's not working for you—it's working for someone else."
The timing matters. We're in the middle of a global AI arms race where China and the US are increasingly viewing AI development through a national security lens. Anthropic likely justified this as some form of security measure or compliance requirement. But here's what they actually did:
- Implemented geographic profiling without user notification
- Created a two-tier trust system where some users get monitored and others don't
- Deployed the capability in production code, not as a one-time security audit
- Only removed it after public exposure, not proactive disclosure
The Implication
This is a precedent-setting moment for agent trust. If Anthropic—the company explicitly marketing itself on safety and ethics—will run covert monitoring on users based on geography, what are the less scrupulous AI companies doing? Every AI assistant you use is making constant decisions about what to log, what to report, and who to tell. You're not the customer when those decisions happen without your knowledge.
For anyone building on Claude or integrating AI agents into workflows: you now have to audit not just what the model does, but what the infrastructure around it does. The agent economy requires agents that work for their users, not against them. Anthropic just showed us how fragile that trust is.