A social-media account affiliated with China Central Television said this weekend that AI company Anthropic PBC has been quietly rewriting its privacy rules in ways that make it easier to hand user data to US intelligence agencies.
The account, known as Yuyuantantian, posted Saturday that Anthropic has changed its privacy policy 13 separate times since 2023. Among those changes, the account said, are provisions allowing the company to move data belonging to users in Canada, Brazil, South Korea and the European Union over to US servers. The post also said Anthropic widened the number of data sources it pulls from, going from three to six between June 2024 and September 2025.
One of the biggest shifts flagged in the post: Anthropic switched its default setting last September so that user conversations are used to train its AI models unless someone opts out, rather than the other way around. For everyday users, that means data that once wasn’t fed into Anthropic’s training pipeline by default now is, unless they actively turn the setting off.
The post also pointed to an Anthropic report in which the company said it shares technical indicators with intelligence agencies after determining that some Chinese companies had been using Anthropic’s models to train their own AI systems without authorization, a practice known as model distillation. Yuyuantantian went further, noting that two of three threat-intelligence job listings Anthropic had posted gave preference to candidates who speak Mandarin or Russian, have intelligence-analysis experience in government or military roles, and hold top-secret US security clearance — details the account used to argue the company’s data-sharing arrangements with US intelligence run deeper than disclosed.
The claims land at a tense moment between Anthropic and Beijing. Neither government has opened a clear path for Anthropic to operate inside China: Chinese regulators haven’t approved the company to do business there, and Anthropic itself has cut off access to its products for China-based users and companies majority-owned by Chinese entities, citing national security concerns. Access to Anthropic’s flagship product, the Claude chatbot, and its coding tool Claude Code has been unavailable in China as a result.
This isn’t the first run-in between Chinese authorities and Anthropic’s software. Earlier this year, China’s Ministry of Industry and Information Technology’s cybersecurity arm flagged versions of Claude Code for what it called a security backdoor, saying the tool could transmit users’ location and identity information to outside servers without their consent. Anthropic disputed that characterization at the time, saying the flagged code was an experimental anti-abuse feature meant to stop unauthorized resellers and prevent companies from copying its models, not a surveillance tool. The company also noted that since its products aren’t available in China to begin with, the warning didn’t affect anyone using its software legally.
For American consumers and businesses using Anthropic’s Claude products, the immediate practical takeaway is a reminder to check privacy settings directly. Anthropic’s own policy lets users opt out of having their conversations used for model training, and the setting is adjustable in the app or web account preferences. Businesses working with Anthropic under commercial agreements are typically covered by separate data-handling terms that don’t feed conversations into training data by default, regardless of the consumer-facing settings.
Anthropic has not yet issued a public response to the Yuyuantantian post as of Sunday. The company has previously said broadly that it discloses its data practices in its published privacy policy and that changes to that policy are communicated to users when they take effect.
JBizNews Desk | San Francisco, Calif.
© JBizNews.com All Rights Reserved. Reproduction or distribution without written permission is prohibited.
