Claude Code 4 min read

Your AI Agent Shouldn’t Hand Out Your Email Address

Key takeaways

  • A User-Agent describes the requesting software; a user’s email address is not required.
  • Establishing whether Claude Code sent an email requires inspecting the request and identifying which component created it.
  • An email in a request header is visible to the receiving server, even over HTTPS.
  • Agents should omit personal identifiers by default when a task does not need them.

Asking an AI agent to check a product’s pricing page sounds routine. But the outgoing request can carry information beyond the page address. What the agent reveals about you deserves attention alongside the answer it brings back.

The request is the evidence

“Claude Code sent my email to a website” is a claim that needs an actual request record behind it. The first questions are concrete: was an email address present, and which server received it?

Then comes attribution. An agent might access the web through a built-in feature, a connected tool, or a script it runs. Each route can involve a different component constructing the request.

A request made during a Claude Code session does not, by itself, establish that Claude Code’s own code inserted the email. A connected tool sending it would still raise a privacy concern. Identifying the responsible component tells us where the fix belongs.

Four details matter:

  • Did the outgoing User-Agent actually contain the user’s email address?
  • Did the request go to the service provider or an unrelated website?
  • Which software version, settings, and connected tools were involved?
  • Did this happen under default settings or because of a customization?

That last distinction determines the potential scope. Evidence from one setup establishes what happened in that setup. A claim affecting all users needs more.

A User-Agent describes the software

A web request includes HTTP headers: fields carrying additional information about the request. One of them, the User-Agent, describes the software making it, typically with a program name and version.

An email address is not a required part of that description. Identifying a program and identifying the person using it are separate jobs.

But the field does not enforce that separation for you. Software can construct a User-Agent string containing an email address and send it to the destination server.

That makes the actual value decisive. A familiar header name tells you little about whether its contents are appropriate.

HTTPS still lets the destination read the request

Suppose an agent includes your email address in a header while fetching a vendor’s pricing page. The vendor’s server can read it. Depending on its logging configuration, it may also retain it.

HTTPS protects the request in transit. The server receiving the request can still read its contents.

The immediate concern is that browsing activity becomes associated with a personal identifier. The pricing-page request now gives the recipient a way to connect that visit to an email address. Repeating the address across requests makes those records easier to link.

This alone does not establish a password leak or an account takeover. Nor does it prove that the recipient stored the email or combined it with other information.

It does mean the recipient received more identifying information than the task may have required. That is a meaningful privacy consequence on its own.

Privacy belongs in the defaults

Fetching public documentation usually requires no disclosure of the user’s account email to the site hosting it. A tool name and version can describe the requesting software.

Some tasks do require identity. Accessing an authenticated service is an obvious example. Those tasks should use the appropriate authentication mechanism and disclose only the information needed for access.

The governing principle should be data minimization: information unrelated to the task stays out of the request.

Making users discover and disable unnecessary identity sharing leaves protection dependent on knowing the setting exists. The default should already do that work.

Products that connect external tools also need to make those tools’ disclosures understandable. An agent’s access to personal information does not automatically justify passing it to every website it visits.

Whether Claude Code sent an email address must be established from the request records. The broader standard is simple: an agent fetching a public page should have a clear reason before handing over your digital business card.

Claude Code AI Agents Privacy

Comments

    Loading comments...