Claude Code users say their AI shared email addresses without asking

A coding assistant may need to visit a website. Does that mean it should introduce you by email address?

A complaint filed in the Claude Code GitHub repository on 5 November alleges that the assistant disclosed a user’s email address in the User-Agent header accompanying web requests. The complainant also blamed the disclosure for scam emails; that causal connection has not been independently established. ¹

A project contributor replied that a built-in privacy safeguard had been introduced in version 2.1.317 and later, enabled by default. Subsequent commenters reported problems despite having the safeguard loaded. One described tests using a deliberately non-resolving address, rather than transmitting the email to a real destination. ¹

Those reports leave the scope and effectiveness of the protection disputed. They do not establish that every Claude Code session exposes personal information. They do raise a concrete question about what an assistant may disclose while trying to finish an otherwise authorised task.

For a business, identity can be part of the information boundary. Permission to research a supplier, inspect a filing or retrieve a document does not automatically settle which employee details may accompany the request.

The awkward part is how ordinary the activity looks. An assistant can deliver the requested research while leaving the user unaware of what it sent along the way. A correct answer is therefore a limited test of whether the task was handled appropriately.

The useful demonstration would show the outbound request as well as the finished answer: which identity was used, what information left the organisation, and whether an effective control prevented unnecessary disclosure.