ANTHROPIC’S recent AI safety report has raised concerns about the privacy of data submitted to its Claude models. The report discusses risks including hacking, surveillance, fraud and the circulation of sensitive data, while indicating that Anthropic has the ability to access complete user logs, including prompts, questions and uploaded files.
The article says this has intensified existing concerns among privacy-conscious enterprise customers, although it does not provide evidence of Anthropic misusing customer data.
Several organisations have reportedly restricted Claude’s use. Nvidia limits its teams to using Claude for low-sensitivity tasks and uses its self-hosted Nemotron models for core internal operations. Palantir has reportedly required an irrevocable zero-data-retention commitment before integrating Claude into customer software, while Booz Allen Hamilton prohibits employees from using it for proprietary cybersecurity work.
The article also says an unnamed utility company abandoned plans to test Claude Fable with core power infrastructure after Anthropic declined to provide such a guarantee.
According to the report, Anthropic introduced a policy in June requiring related prompts and outputs to be retained for 30 days by default to help detect complex cross-session attacks. Content flagged by a safety classifier may be stored for longer. Anthropic says enterprise data is not used to train models without explicit authorisation, but the article says regulated customers remain concerned about persistent retention and possible access to proprietary conversations.