Live Updates > Live Update Details
2026-09-19 15:08:09
[Yu Yuan Tan Tian: Anthropic Hands Over Global User Data to US Intelligence Agencies After 13 Revisions to User Privacy Agreement in 3 Years] On September 12, Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and xAI founder Elon Musk reached a rare public consensus late at night in the US, believing that AI development needs to slow down. Ironically, Amodei, who champions AI security, acts in stark contrast. On September 14, foreign media revealed that Nvidia and several other US companies have begun restricting or halting the use of Anthropic's cutting-edge models. One of the triggers for US companies ceasing to use large US models is Anthropic's unilateral modification of its data retention agreement: mandating the retention of user interaction data for 30 days for security reviews, with users having no right to refuse. This is not the first time it has modified its user data agreement. Tan Zhu reviewed the 13 revisions of Anthropic's Privacy Policy since its initial release in 2023 and compared them with the ISO/IEC 27701 privacy information management system standard, finding that Anthropic's user privacy data security risks are increasing. To become an intelligence center, data alone is not enough; you also need people. In July 2026, Anthropic posted three "Threat Intelligence Manager" positions on its website. Two of these positions prioritized candidates with the following qualifications: fluency in Mandarin or Russian, experience in intelligence analysis in government or military environments, and a U.S. "Top Secret Security Clearance." These positions were responsible for investigating whether user interaction data contained information about foreign government-backed propaganda manipulation or model "distillation." On September 10th, an Anthropic threat intelligence report, highly consistent with the job descriptions in the July job postings, was released, analyzing approximately 200 million interactions between users and the Claude model suspected of being "distilled." The strategy involved first collecting data worldwide to define "threats," then recruiting "insiders" to analyze the data, and finally adapting the company's security standards to U.S. standards. When American standards are pushed to the world, the "risks of AI development" mentioned by Amodei will naturally cease to exist.
