A developer using Claude Code discovered this week that the tool was inserting his registered email address into the User-Agent header of every HTTP request it made during web searches. The data went to every website the tool visited, without explicit consent or any obvious engineering reason.
GitHub issue #107757 documents the finding, and the Hacker News discussion quickly passed 300 points. The concern is not about malice — it is about the gap between what AI coding agents do and what their users expect.
A preventable oversight
The User-Agent string identifies the client making a request. Including a personal email there means any server-side log, analytics tool, or third-party script embedded on a visited page could harvest it. This is a configuration mistake, not a sophisticated exploit. Claude Code version 2.1.318 on macOS exhibited the behaviour, and a simple fix — stripping identifying data from the header — addresses it immediately.
Anthropic has not yet publicly commented on the disclosure, and the fix timeline remains unclear.
What good looks like
Developer tools with network access should treat privacy boundaries as core engineering constraints, not afterthoughts. The right approach is explicit: tell users what data the tool sends to remote servers, provide a way to inspect it, and never include personal identifiers in headers unless the user opts in with clear understanding of where it goes.
This incident is small in scale, but it reflects a pattern. As AI coding assistants become more autonomous — searching the web, calling APIs, executing commands — the surface area for inadvertent data exposure grows. Trust in these tools depends on whether their makers earn it, request by request.