Kuaikeji, September 19 - Anthropic, one of the leading cutting-edge large model companies in the US, has recently been exposed for directly submitting user interaction data from around the world to US intelligence agencies, a move that has raised widespread concerns across the global tech industry.

According to authoritative overseas media, several US-based tech companies, including NVIDIA, have begun restricting or completely discontinuing their use of Anthropic's cutting-edge large model services. The primary catalyst for US tech companies abandoning this leading domestic large model is Anthropic's unilateral modification of its data retention agreement, mandating that all user interaction data be retained for 30 days for security reviews, without providing users any option to refuse. This effectively transfers complete control over personal data.

This is not the first time Anthropic has made protocol adjustments that compromise user data security.

Cybersecurity agencies have compared all 13 revisions to Anthropic's privacy policy since its initial release in 2023 with the internationally recognized privacy information management system standard ISO/IEC 27701. They found that nearly every update has loosened the company's permissions for accessing user data, continuously increasing the security risks to ordinary users' privacy data.

Building a large model data processing center to serve intelligence systems requires more than just massive data reserves; the gap in professional analysts must also be quickly filled.

In July 2026, Anthropic posted three job openings for threat intelligence managers on its official website. Two of these positions explicitly prioritized candidates fluent in Mandarin or Russian, with experience conducting intelligence analysis in government or military settings, and holding top-secret US security clearances.

The core responsibilities of these two roles include identifying foreign government-led opinion manipulation in user interaction data and illegal technical distillation theft targeting the Claude model.

By September 10, an Anthropic threat intelligence report, closely aligned with the job descriptions posted in July, was officially released. This report analyzed approximately 200 million instances of global user interactions with the Claude model suspected of involving so-called technical distillation, with all ordinary users' daily behaviors included in the indiscriminate screening.

The logic behind Anthropic's operational path is now clear: first, collect massive amounts of interaction data from users worldwide and self-define various so-called security threats; then, recruit personnel with official security clearances to conduct targeted data analysis; finally, gradually replace internal corporate security standards with those fully compliant with US intelligence system requirements.

When this US-led framework is further promoted globally, the so-called AI development risks repeatedly emphasized by its leaders will naturally become tools serving US interests.