B[Yuyuan Tantian: The user privacy agreement has been revised 13 times in 3 years, and Anthropic handed over global user data to US intelligence agencies] Silicon Valley in the United States is setting off a fierce debate around the speed and security of AI development. 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 United States that AI development needs to slow down. Although Altman only responded to one of Amodei's suggestions - embedding third-party evaluators; Musk only said "Dario is right" in this regard, this topic has quickly divided Silicon Valley into two camps recently.
On the other side, Nvidia CEO Jensen Huang, Meta CEO Mark Zuckerberg and others have made it clear that companies should "run as fast as possible" and can stop on their own if they feel it is unsafe. There is no need to "press the brakes" uniformly.
On the surface, this debate is a battle over technical lines, but in reality the dispute is over who has the right to define AI safety standards. Paradoxically, Amodei, who claims AI is safe, actually behaves contrary to this.
On September 14, foreign media revealed that many American companies such as NVIDIA have begun to restrict or stop using Anthropic’s cutting-edge models.
American companies began to no longer use large American models. One of the triggers was Anthropic's unilateral modification of the data retention agreement: it is mandatory to retain user interaction data for 30 days for security review, and users have no right to refuse.
This is not the first time it has revised its user data agreement.
Master Tan sorted out the 13 version revisions since Anthropic released the first version of the "Privacy Policy" in 2023, and compared it with the privacy information management system standard ISO/IEC 27701, and found that Anthropic users' privacy data security risks are getting higher and higher.
Most mainstream AI companies at home and abroad have stated in their privacy policies that they will cooperate with law enforcement agencies’ data retrieval requirements in accordance with the law. But what is different about Anthropic is that, in addition, its privacy policy stipulates that as long as it feels necessary, it can share user data with US intelligence agencies without going through legal procedures. The judgment and standards are completely defined by the enterprise.
Why does Anthropic do this? Analyzing the content and timing of each privacy policy modification, we can find three steps for Anthropic to become an "intelligence center."
First, transfer overseas user data to the United States.
Starting from May 2024, Anthropic mentioned in five versions of its privacy policy that relevant data of users in Canada, Brazil, South Korea, the European Union and other places will be transferred to the United States in accordance with the terms by registering subsidiaries or supplementary regulations. The current policy for Canadian users even states that U.S. data protection standards may be looser. This means that as soon as the user's data reaches the United States, it is no longer protected by the standards of the country where the user is located.
Second, expand the sources of user data.
Between June 2024 and September 2025 alone, the number of user data sources available to Anthropic increased from 3 to 6, including user conversations, self-generated data, and user identity data newly added this year. These data can be recovered and used to train Anthropic's large model. The larger the amount of data and the more types, the stronger the ability of the trained large model may be. To this end, Anthropic also simultaneously revised the regulations related to large model training - from not being used by default in 2024 to being used by default in September 2025. Even if the user refuses to provide data, Anthropic may flag it on the grounds of "security" and still use it for large model training.
Third, provide intelligence to U.S. intelligence agencies in exchange for benefits.
On February 23, 2026, Anthropic released a report saying that it found that Chinese companies were involved in so-called "distillation" activities, and stated in the report that it "is sharing technical indicators with relevant intelligence agencies." Since then, Anthropic has continued to provide additional relevant intelligence: On June 10, 2026, it submitted a letter to the U.S. Senate, stating that by analyzing 28.8 million user conversation data and 25,000 user accounts, it had obtained new information on Chinese companies’ participation in so-called “distillation” activities.
To become an intelligence center, it is not enough to have data, you also need people. In July 2026, Anthropic’s official website announced three “Threat Intelligence Manager” positions. Two of the positions give priority to the following conditions: Fluency in languages such as Mandarin or Russian, experience in conducting intelligence analysis in a government or military environment, and holding a "Top Secret Security Clearance" in the United States. They are respectively responsible for investigating whether there is public opinion manipulation or model "distillation" supported by foreign governments in user interaction data.
On September 10, an Anthropic threat intelligence report was released that was highly consistent with the responsibilities of the July recruitment position. It analyzed approximately 200 million alleged so-called "distilled" interaction data between users and Claude models.
On September 12, Anthropic CEO Amodei cited this report and proposed three plans to "slow down" the development of AI.
First, collect data from all over the world to define "threats", then recruit "our own people" to analyze the data, and then turn the company's security standards into American standards. When American standards are introduced to the world, the "AI development risks" mentioned by Amodei will naturally no longer exist.
At this time, looking back at the "technical consensus" reached by the three giants, it is essentially a problem of the US AI industry itself. What they want to slow down is the speed at which others are catching up; what they want to speed up is their own collection of data, mining of intelligence, and definition of rules and security. (Yuyuan Tantian)
— All loaded —
