Your coding agent leaked your email to websites this week, and the fix shipped the same day. That is the story everyone told.
It is the smaller story.
THE OLDER REPORT
On July 25, a developer decompiled the shipped Claude Code bundle and filed issue #81138: the signed-in user's account email is read from a local settings file and injected into the model's system prompt on every session. The quoted bundle code is short and plain. The line it produces reads like a stage direction: "The user's email address is" — followed by yours.
That issue is still open tonight. Anthropic staff have said the injection was intentional: the address sits in context so the model can identify you, instead of inventing fake author tags. An instruction in the same context tells the model not to send that address to unrelated services unless you ask.
Read that twice. The guardrail for a piece of personal data placed inside the prompt is another line in the same prompt — a request, addressed to the component most likely to be talked out of it.
PIPE TWO
On November 5, issue #107757 reported the address leaving in the User-Agent header of web requests — the reporter on version 2.1.318, reactions in the hundreds within a day, an inbox filling with scam mail as the receipt.
Anthropic shipped a built-in mod, cc-plugin-privacy, in 2.1.317: it refuses shell calls that would send the account email out in a request. Its own README lists what it does not cover — the built-in web tools among them — and users on the newer build report cases where the block never triggered. Staff have since said the real fix is scheduled: the email comes out of the model's context entirely in a release due the following week. Version 2.1.319, out November 6, mentions neither the email nor the mod.
THE LESSON IS NOT ABOUT ONE VENDOR
Your system prompt is not a private notebook. It is a document that leaves the building: into logs, into traces, into subagent contexts, into every tool call the model narrates. Anything written into it should be treated as published-to-the-machine — readable by every component downstream, forever, including the ones having a bad day.
Identity belongs in the harness, in auth headers the model never sees. The moment an identifier is placed in model-visible context "so it knows who you are," you have made a prompt do a credential's job — and prompts are the one surface in this stack designed to be persuaded.
DO THIS BEFORE MONDAY
Read your own system prompt the way an auditor would. Dump the full context of one session. Anything personal in it is a distribution decision someone made for you.
Grep your agent's config for your own identifiers. Email, name, account IDs. If the model can read it, assume a tool call can carry it.
Prefer harness-set identity. Headers, tokens and author fields set by the wrapper — never by the model's memory.
Watch the vendor's context-removal release. When it lands, verify it the way the reporters did: a local listener, one fetch, read what arrived. Trust is a test you can run.
Sources
GitHub issues #81138 (system-prompt injection, filed Jul 25, open) and #107757 (User-Agent leak, filed Nov 5) on anthropics/claude-code, including the staff replies and the cc-plugin-privacy README; Claude Code changelog for 2.1.317 and 2.1.319; independent write-ups of both issues (dev.to, madrobot.blog, thetechlane.in), November 2026.
Question for you: open one session of the agent you use most and dump its full context — what's the most personal line in it, and did you put it there?
Disclaimer
This post describes my personal experience with my own systems and accounts, in my own words. It is a first-person account of what I observed and what I do in my own work — not a statement of any company’s intentions, policies, or internal decisions. Where I do not know something, the post says so.
I am not affiliated with, endorsed by, or speaking for any company, product, or vendor named in this post. All product and company names are trademarks of their respective owners and are mentioned only to describe the experience accurately.
Any error messages, documentation excerpts, or figures quoted here are reproduced from my own records or from the public sources linked in the post, believed accurate at the time of writing. Product behavior changes over time; verify against current official documentation before relying on any detail.
Nothing in this post is professional, financial, legal, or security advice for your specific situation. Examples are illustrative — validate in your own environment before running anything in production.
Corrections are welcome. If any company named here publishes an official statement, or if any detail is wrong, I will update this post and note the correction. The goal is an accurate public record, not a verdict.
© 2026 Honey Kiran. All rights reserved.
Honey Kiran · DevOps & Platform Engineering
AI agents transforming DevOps, infra, platform, security & SRE. The shift nobody works anymore: deep dives, 3AM Files, and OpsBites.
Examples are illustrative — validate in your own environment before running in production.

Top comments (0)