On November 5 a Hacker News post started a very bad week for Claude Code. The author was receiving dozens of phishing emails a day that addressed him by his real name, sent from domains he had never visited. Following the request logs backwards led to the terminal tool itself: while fetching pages from the web, it was folding the session name and the account email into the User-Agent header.
How the email got into the context
The origin is almost innocent. Users had complained that Claude invented names when writing author annotation tags, because the model had no account information to work with. The team's fix was to read the login email from local config and splice it into the system context at startup.
The field served no product function. It was documented nowhere and appeared in no settings screen. But from that moment it was something the model could see — and anything the model can see, it can use.
The leak happened at the 403
The SEC's filing API, Wikipedia, crates.io and the EVE Online endpoints all ask scripts to leave a contact address in the request headers, and return 403 if one is missing.
Claude's task was to fetch the filing. The 403 was an obstacle, the context contained an email, so it appended that email to the User-Agent or wrote it straight into a shell command. It later explained the reasoning: the site needed a contact so the IP would not be blocked again.
Sub-agents made it worse. A developer posted server logs in which a sub-agent fetching external documentation wrote the user's real name, team name and internal GitHub repository path into the request headers. This was not one slip. Every link in the chain was trying to be helpful.
Four months of tickets, two non-fixes
The community did report it. From April 2026, at least four tickets (50702 and 53231 among them) said the account email was being forced into context with no way to disable it. A bot closed all of them as stale or duplicate. In mid-July someone filed 78431 after curl leaked the address; in August the story reached Hacker News; in September the ticket was closed again.
On July 25 the developer skflowne decompiled the v2.1.220 package and filed 81138 with receipts: the client read the login email from local config at every startup and built it into the prompt, and the field had no purpose beyond two boolean telemetry events.
The official repairs came in stages. v2.1.246 in late August kept the field and appended a sentence asking the model to use it only for identification and never send it to unrelated services. v2.1.289 in late October moved that sentence into the system reminder attached to every tool call, repeating it on every request.
Natural language is a weak constraint on a probabilistic model. On October 20 a user caught Opus 5.5 putting the account email into the headers of two shell commands. Confronted, the model conceded that the context held the address and that instructions forbade sending it, and that it had sent it anyway to leave a contact for the target site.
By early November the argument had gone public. Engineer mpoteat said v2.1.317 shipped a default-on built-in plugin that blocks requests containing the user email with deterministic rules — an interceptor rather than an instruction. Explaining that "controlling model behaviour matters and is genuinely hard", he ended with "haha"; the screenshot travelled, the joke landed badly, and he deleted it.
Hours later Lydia Hallie added the part that mattered more: the plugin cannot stop Claude from writing its own script to make the request, and the change that removes the email from context would arrive in the regular release early the following week. As of November 9 the public build was v2.1.319, upgradable with claude update.
The IS_DEMO detail is the bitter one
Claude Code exposes an IS_DEMO environment variable, recommended for streaming or screen recording, which hides the email and organisation name from the status bar. The decompiled package shows all seven checks gate on-screen text only. The email in the model's context was untouched.
Demo mode protects the audience's eyes. It does not stop the model from handing the address over.
Four things to do now
First, upgrade. Run claude update and confirm you are on v2.1.319 or later; the next release removes the context field, so take it when it lands.
Second, audit your history. If you have worked against any API that asks for a contact address, check the User-Agent strings and shell history from that period for your email, your real name or repository paths. If they are there, the address is already in someone else's log.
Third, narrow network and command permissions. Use deny rules to block curl and wget calls you do not need, or route network actions through an approval list. For an agent, "may it make this request" is far easier to control than "what does it put in the request".
Fourth, separate identity from account. Subscribe with a dedicated alias and pin an anonymous commit identity in your repositories (git config --global user.email). Do not let the login address also be the commit address, the billing address and the public contact address.
This is not the end of the bug
Removing the field solves half of it. The sites that demand a contact address still exist. When the next 403 arrives, will Claude reach for a different field to finish the task — the username, the team name, the project name? Nobody has promised otherwise.
For agent products, the boundary between autonomy and controllability was not drawn here. Next time it will not be called an email leak. It will be called something else.