Claude Code accused of leaking users' email addresses to websites

A problem known since July

The issue became one of the top topics on Hacker News because it is not the first of its kind. Similar reports were opened in July and again in October. One commenter, phase3dev, put it this way: "Don't underestimate the intelligence of your users."

According to issue #81138, Claude Code reads the account address from a local configuration file and adds it to the context the model sees at the start of every session. When the model contacts a site such as Wikipedia, which asks automated programs to identify themselves with a contact, it has sometimes copied that address into its request.

A fix that only partly works

Anthropic has not directly acknowledged the problem. A contributor to the repository did reply, however, that a protection module built into Claude Code and enabled by default has blocked this behavior since version 2.1.317. A fix has been available since 3 November.

That has not calmed the criticism. One developer tested the module with 23 commands containing a fake address. By his account, it does block sending the address in a User-Agent header, but it lets through eight other ways of transmitting it: in another header, in a login identifier, or when the command fetches the address itself at the last moment.

For developers, the practical risk is that a personal address ends up in the logs of sites the agent contacts, with no obvious warning. Whether Anthropic will extend the protection to the remaining routes has not been stated.