Claude Code felfedte a felhasználó e-mail címét bizonyos webhelyek meglátogatásakor

A Claude Code AI agent has identified a privacy issue that led to the leakage of email address information when accessing certain third-party websites. The email was included in the User-Agent HTTP header without the user’s consent and without the ability to disable such behavior. Substitution occurred for specific sites such as Wikipedia and the SEC (Securities and Exchange Commission), which do not accept requests from bots that do not specify the owner’s email. According to a comment from a representative of Anthropic, this activity was not a purposely added feature but rather an initiative of the AI model responding to the formatting requirements for bot requests and the recommendations of RFC-9110, which stipulates that a contact email should be provided for automated inquiries.

Information about the email linked to the user’s account was transmitted to the AI model in Claude Code so that it could consider context when processing tasks related to the user and their account. Without such context, the model hallucinated and used fabricated information, for example, in the tags ‘@author’. A side effect was that the AI model started using the email not only for expected purposes but also while performing ancillary tasks, such as inserting it into the User-Agent field when generating command strings for the curl utility.

To block the leakage of email in external requests
a cc-plugin-privacy plugin has been prepared, which is included in the standard set of plugins and activated by default in the claude-code 2.1.317 release. The plugin asks the user for confirmation before using web tools or generating commands in curl style that mention the email. The plugin does not address all possible leakage paths, for example, it does not help in situations where Claude generates and runs a script itself.
To fully block the issue, the next release of Claude Code, which is planned to be released in a few days, will remove the email from the context passed to the AI model.

Forrás: opennet.ru