Claude Code leaks user email addresses without authorization; Anthropic admits fix still has vulnerabilities.

Beating AI News Flash, November 5 — A developer complained on GitHub that Claude Code leaked his email address, and he has recently been receiving dozens of scam emails every day.

Anthropic engineer Lydia Hallie explained that Claude previously often filled in author names arbitrarily when writing code. To solve this problem, the team added users' login email addresses to the model context.

When accessing certain websites, Claude writes its own network commands. Some websites require automated programs to leave a contact email, and Claude fills in the user's real email address without authorization and sends it along with the request. Even system instructions explicitly prohibiting the disclosure of personal information failed to fully prevent this behavior.

This issue was first reported as early as July this year. Since then, developers have successively discovered that Claude Code sent email addresses to websites including Wikipedia and the U.S. Securities and Exchange Commission. One user found 15 similar records across 4 projects, with some requests initiated by sub-agents on their own.

Anthropic has added a privacy interception feature in version 2.1.317 that prompts the user before a tool call is about to send the account email address. However, if Claude first writes the email address into a Python script and then runs the script, it may bypass the check.

Anthropic acknowledges that the current solution is not thorough and plans to directly remove login email addresses from the model context in the next version.