Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Claude Code sent your email to websites

Written November 8, 2026

Claude Code put your account email in the model's context, and Claude copied it into User-Agent headers. 2.1.317 asks first; the full fix comes next.

What happened

A bug report opened on November 5, at 22:13 UTC, on Claude Code 2.1.318, says the tool divulged the reporter's email address to third parties in the User-Agent header of requests it sent while doing web searches. The report carries the bug and area:security labels, and the reporter says the leak is why they now receive dozens of scam emails a day. That last part is the reporter's account, not a measured fact, but the leak itself was confirmed by the Claude Code team the next day.

The cause is ordinary, which is what makes it worth understanding. Users had reported Claude inventing names for things like @author tags, so the team added the signed-in account's email to the model's context. When Claude then wrote a command itself, such as a curl call, it sometimes put that address in the User-Agent header as a contact address. Some sites ask automated clients for a contact, and Claude supplied the one it had. Lydia Hallie of the Claude Code team explained this on Hacker News on November 6 and apologized for the team, writing that this should not have happened. An earlier issue, opened on July 25, had already reported that the account email sat in the model's context, as of version 2.1.220.

The two-step fix

The fix comes in two steps, and only the first has shipped. Version 2.1.317 adds a built-in mod, cc-plugin-privacy, on by default, that asks before Claude runs a web tool call or a curl-style command containing your email. A Claude Code team member confirmed in the issue thread that the mod is enabled by default in 2.1.317 and later, and that /plugins turns it off. The second step, removing the email from the model's context entirely, comes in the next release, which the team expects early next week.

Two version facts matter for anyone checking. The stable channel was still on 2.1.311 on November 6, so it does not have the check yet. And the check does not catch every case: the team says it does not prevent a script Claude writes and then runs on its own. Version 2.1.319, the latest on November 7, does not mention the account email in its changelog. The gap is real but bounded: the check covers the web tool calls and curl-style commands the team could see, and the full fix is the removal of the data itself.

What to do now

  1. Check your version. Run claude -v. Anything below 2.1.317 does not have the check.
  2. Update, on the right channel. Run claude update. It follows your release channel, so on the stable channel it keeps you on the stable version. Switch first: open /config and change the auto-update channel, or set "autoUpdatesChannel": "latest" in settings.json. Homebrew picks the channel by cask: claude-code tracks stable and claude-code@latest tracks latest. The apt, dnf, and apk repositories pick it by repository.
  3. Leave the mod on. Check in /plugins that cc-plugin-privacy is still enabled. It ships on by default, and turning it off removes the only check between Claude and the network.
  4. Make network commands ask. The mod does not catch a script Claude writes and runs, so read the headers any such script sets before you approve it. Ask rules make curl and wget prompt every time:
    • "permissions": { "ask": ["Bash(curl *)", "Bash(wget *)"] }
    A Bash rule matches the command form Claude usually writes, not /usr/bin/curl or a shell-wrapped call. For a limit that must hold, pair a deny rule with the sandbox network allowlist.
  5. Look back through recent sessions. Claude Code keeps session transcripts in plain text under ~/.claude/projects/ for 30 days by default. Search them for your address inside commands Claude ran, such as curl calls. Because the address was in the session's context, a plain search can also match lines that are not requests. A match in a command shows Claude wrote it, not that you approved the request.

Why an instruction did not stop it

Telling the model not to share the email was not enough. Version 2.1.234 already told Claude to use the account email only to identify you, and not to send it to unrelated services unless you ask, and the behaviour continued. Anthropic's permissions docs say the same about your own instructions: guidance in CLAUDE.md shapes what Claude tries but does not enforce a boundary. What holds is a check that runs outside the model, like the 2.1.317 mod, a permission rule, or the sandbox, or not giving the model the data at all, which is the next release's fix. The general lesson for anyone running agents is that anything in an agent's context can end up in what it writes, so the limits that matter belong in settings, not in prose.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Claude Code 2.1.319: unattended runs stop stalling · Audit a repository's git config before any agent opens it · Check your Codex and Claude Code config for relay tampering