Rapid Response • September 18, 2026
A newly published Anthropic threat-intelligence report adds a separate but relevant data-governance development to the questions raised by Free China Movement’s DeepSeek black-box audit.
Anthropic reported that it observed DeepSeek silently relaying selected user exchanges to Claude without informing DeepSeek customers. According to Anthropic, some relayed exchanges contained sensitive information, including internal technical material and, in one case, live database credentials. Anthropic also described exchanges related to a Chinese public-security surveillance system.
This is not evidence of the internal mechanism behind politically sensitive refusals documented in Free China Movement’s separate audit. It is, however, a new external development relevant to the audit’s sensitive-information protocol: researchers should assume that model routing and data handling may involve layers not visible from the user interface and should avoid placing unnecessary confidential or identifying data into black-box tests.
The Free China Movement audit therefore keeps separate: training data, retrieval access, moderation/policy layers, model routing and final output. Claims about one layer should not be inferred from another without evidence.
Read the DeepSeek audit: https://freechina.live/english/deepseek-black-box-censorship-audit-politically-sensitive-response-restrictions-2026/
External source: Anthropic, Countering misuse of AI: September 2026.