[Claude Code Leaks User Emails Without Authorization; Anthropic Admits Fix Still Has Vulnerabilities] Beating AI News, November 5th – A developer complained on GitHub that Claude Code leaked his email address, and he has been receiving dozens of scam emails daily recently. Anthropic engineer Lydia Hallie explained that Claude frequently entered incorrect author names when writing code. To address this issue, the team added the user's login email address to the model context. When visiting certain websites, Claude would write its own network commands. Some websites required the automation program to leave a contact email address, and Claude would enter the user's real email address without authorization and send it along with the request. Even though system instructions explicitly prohibited the disclosure of personal information, this behavior was not completely prevented. This issue was first reported in July of this year. Since then, developers have discovered that Claude Code sent emails to websites such as Wikipedia and the U.S. Securities and Exchange Commission. One user also found 15 similar records in 4 projects, some of which were initiated by sub-agents. Anthropic added privacy filtering in version 2.1.317, prompting the user before sending the email address when a tool is about to send it. However, if Claude writes the email address into a Python script before running it, it might bypass the check. Anthropic acknowledges that the current solution is not thorough and plans to remove the login email address from the model context in the next version. [Original Link]