Anthropic’s new Claude feature can leak data—users told to “monitor chats closely”

On Tuesday, Anthropic unveiled a new functionality for its Claude AI assistant that allows users to create Excel spreadsheets, PowerPoint presentations, and other documents directly through its web interface and desktop application. While the feature promises to enhance usability, the company has issued a warning in its support documentation, stating that it could potentially compromise user data. Dubbed "Upgraded file creation and analysis," this new tool resembles ChatGPT's Code Interpreter and acts as an enhanced version of Anthropic's previous analysis capabilities. Currently available in preview for users on Max, Team, and Enterprise plans, Pro users can expect access in the coming weeks. The security risks stem from the feature’s ability to operate within a sandbox computing environment, which allows Claude to download software packages and execute code to generate files. Anthropic has cautioned that this capability may expose user data, advising users to "monitor chats closely when using this feature." According to the documentation, a malicious actor could potentially exploit this feature by inserting covert instructions through external files or websites, enabling Claude to access sensitive data from connected knowledge sources. This manipulation could lead to data being leaked via the AI's sandbox environment. This vulnerability is linked to prompt injection attacks, where hidden commands can influence the AI's behavior, a security concern first identified by researchers in 2022. Anthropic has acknowledged these vulnerabilities, stating they were discovered during extensive security testing prior to the feature's rollout. The company recommends that users vigilantly observe Claude's actions while using this feature and intervene if they notice any unexpected data handling. However, this places the onus of security on users, despite the feature being marketed as a seamless, automated solution.