Claude Code users report their email addresses were sent to websites in User-Agent headers
A GitHub bug report says the coding tool put signed-in users' email addresses in web requests. A commenter who says they work on the tool says a fix is coming.
Users of Claude Code, Anthropic's coding tool, say it sent their account email addresses to third-party websites in the User-Agent header of web requests. That is according to a bug report on GitHub and a linked Hacker News discussion, the only source for this story.
The person who opened the GitHub issue said the tool exposed their email while running web searches without their permission, and that they now get dozens of scam emails a day. Other commenters described sudden floods of spam. One said their logs confirmed their email had been sent to other domains.
On Hacker News, a commenter posting as lydia_hallie, who wrote "I work on Claude Code," explained the cause. Users had complained that the model made up names for things like author tags, so the account email was added to the model's context for signed-in users. When the model then wrote its own commands, such as a curl call, it sometimes put that email in the User-Agent as a contact address. The commenter said some sites ask automated agents for a contact.
“I'm sorry we let this happen.”lydia_hallie, on Hacker News
The commenter said the email would be taken out of the model's context in the next release, expected early next week.
This matters because users say a tool they run on their own machines handed personal contact details to unknown websites without asking them. Several commenters said privacy protection should be built in, not offered as a plugin.
For now, a built-in plugin called cc-plugin-privacy is offered as a stopgap in version 2.1.317 and later and is on by default, according to the thread. The descriptions of it differ. A reply on GitHub said it was meant to "deterministically block this behavior." The Hacker News comment said it asks the user before a web tool call or curl-style command containing the email is run, and that it "still doesn't prevent all cases," such as when the model writes and runs a script itself. That comment also said the stable release channel was still on version 2.1.311.
Some questions are still open. One user on version 2.1.318 said the plugin was not present at all. Another, on macOS, said the plugin showed as loaded and enabled but did not refuse a request it was documented to refuse, and asked whether the shipped build differs from the published source. The thread does not answer this.
It is also not known how many users were affected or which sites received their addresses. One commenter said the problem had been reported before, including in a GitHub issue opened five months earlier. That claim could not be checked against the source. The source also does not confirm the identity of the people posting.