What the Context Can Send

A bug report says Claude Code sent an account email to websites. The model sometimes wrote it into commands it composed; the check that shipped covers one shape, and the proxy log I've been wiring up can't see inside TLS.

Issue 40 - Nov. 10, 2026
By -top:1px solid #d4d4d8;font-size:0;line-height:0"> 

The News Voted Otherwise

I closed last week's issue by promising we'd spend this one on the logs an agent can reach; the news voted otherwise, and it voted loudly. A Hacker News story headlined "Claude Code sends your email address to websites" went up Nov. 5 and stood at 1,608 points by Nov. 9. It links to GitHub issue #107757 on anthropics/claude-code, "Bug Report: Privacy violation in User-Agent header," opened the same day by a user named taylor-edwards on version 2.1.318 on macOS and still open Nov. 9. The reporter's claim is blunt: "I receive dozens of scam emails per day because claude code divulged my email to arbitrary 3rd parties in the User-Agent request header while conducting web searches without my permission."

The headline needs one correction before anything else. Nothing in the tool stamps a fixed email header on your traffic. A commenter posting as lydia_hallie wrote on Hacker News Nov. 6: "I work on Claude Code. We're removing the account email address from the context Claude Code gives the model." By that commenter's account, the address sat in the model's context for signed-in users, and commands the model composed on its own (curl calls, for instance) sometimes used it as a User-Agent contact. The behavior is model-driven and intermittent; the model had the value in front of it, and sometimes it wrote the value down.

My position comes from someone whose own workflow runs on this tool, so weigh it accordingly. Whatever sits in the model's context is on the egress path. The check Anthropic shipped covers one shape, by its own README; the controls that last live upstream and downstream of the model, which means keeping out of context what you wouldn't send and reading the record for what left. The twist for this series comes at my expense, because the proxy log that the meter from #37, the countersign from #38, and the tripwire from #39 all hang on can't see inside TLS.

 

FOR FURTHER READING

 

The Claim and the Explanation

Before anyone grabs a pitchfork, the report has real holes. The issue shows no packet capture, no named destinations, and no reproduction steps. The causal claim, that this behavior produced the scam mail, is unproven; GIGAZINE on Nov. 6 and Kernelpost on Nov. 7 both flagged it as unconfirmed, and the flag stands here too (an inbox full of scam email has no shortage of possible parents).

The explanation from the Claude Code side holds up better, and it's oddly mundane. Users were seeing invented names in places like @author tags, and the remedy was to hand the model the real thing: "So we added the account email in the model's context for signed-in users." Some sites ask agents for a contact, so commands Claude composed itself sometimes filled that slot with the address it was holding. The commenter apologized, said this should've been caught sooner, and put removal on a short clock: "This update will hit Claude Code users in the next release, which should arrive early next week."

The newest version in the changelog, as of a Nov. 9 read, is 2.1.319, and no entry says the address left the context; removal is announced and, as far as the changelog shows, unshipped. I won't write "fixed" until it's on that page. (WebFetch's own default User-Agent is "Claude-User," per a changelog entry, so my reading is that the reporter's "web searches" case probably involved a command the model wrote. That's inference, labeled as such.)

 

A Ticket, a Bot, and Four Months

The public record starts July 17, when lgualtieri75 filed issue #78431 against v2.1.212 reporting the same User-Agent behavior; a stale bot closed it as "not planned" Sept. 11 (the bot, to its credit, has never once pretended to read a ticket). The best evidence in that thread came from mt-empty, who posted a session log Aug. 11 showing a curl call with a User-Agent carrying an email, destination crates.io, five times in one hour. That finding came from reading session logs and involved no network capture, a detail that comes back later.

Another commenter pointed to the crates.io policy, which recommends contact information in the User-Agent, so the model may have been following it; plausible, and unproven. A different commenter, teidesu, couldn't reproduce the behavior, which is about what I'd expect from something a model decides to do on the fly. The only destination documented in a posted log is crates.io, and no other gets named here.

Issue #81138, filed July 25 against 2.1.220, supplied the mechanism. Its author, skflowne, decompiled the bundle and reported that Claude Code injects the OAuth account email into every session as a # userEmail block; commenters later reported the same on later versions, and the Hacker News comment confirms the address was in the model's context. That's a public ticket in July, a public teardown later that month, a public log in August, and the thread that made it news Nov. 5, roughly four months after the first report.

The standards had a line on this before any of those tickets. RFC 9110, published June 2022, says in section 10.1.2: "A user agent SHOULD NOT send a From header field without explicit configuration by the user." These reports concern User-Agent, so that line applies by analogy only; the analogy's still close, because an email address in a request header is a From field wearing a mustache. "A user agent SHOULD NOT generate a User-Agent header field containing needlessly fine-grained detail" is section 10.1.5, and that one needs no analogy. An account email is about as fine-grained as detail gets.

 

A Courtesy and One Shape

Anthropic has earned some credit here, and I'll pay it before the caveat. It responded, apologized, committed to the removal, and shipped a check; the 2.1.317 changelog entry, dated Nov. 3, reads "Added a check that asks before a web or shell tool call sends your account email address to the network" and predates the thread that made this news. A contributor named poteat wrote on the issue Nov. 5 that a built-in mod called cc-plugin-privacy blocks this in v2.1.317 and later and is enabled by default.

The vendor shipped a guard, turned the guard on, and promised the removal, so the matter's closed. Right? The mod's README answers that one in its own voice: "This is a courtesy, not a boundary, and it covers one shape." Per that README it matches plain words, so it can false-refuse something as harmless as curl -user (anyone who's worked with an overeager linter knows this coworker); and it misses an address in a URL, a request body, curl -u, a From: header, or an address computed at run time. WebFetch, WebSearch, and MCP tools aren't checked by it either. That's a smaller claim than the changelog's "web or shell tool call" suggests; my reading is that the two describe the same check, and I'd trust the README over the one-liner. (The reporter's version, 2.1.318, is one release past the check, and with no reproduction steps on the issue, nobody can say whether it was in play.)

My criticism comes in two parts, and the fact that I rely on this tool is a reason for more candor. Roughly four months is a long time for a report like this to sit in public, and the commenter said as much about catching it sooner. A check also stops being a boundary at the first string it can't match; the commenter conceded the gap in plain language: "But it still doesn't prevent all cases, like when Claude writes a script and runs on its own." Community workarounds are circulating, and nothing in Anthropic's docs confirms them, so none of them gets a recommendation from me.

No confirmed report of this specific bug at any other vendor has reached me, and I won't imply one. My opinion is that the class is general: any agent that holds a value in context can write that value into a command. An account email is the case we got the receipts on; nothing about the mechanism cares whether the value is an address, a customer name in a CLAUDE.md, or an internal hostname.

 

What the Proxy Log Can't See

I owe you a correction of my own. Three issues running, I've told you to wire your controls to a Squid log. Squid's logformat can record one named request header with %{User-Agent}>h, which sounds like the whole answer until TLS enters the room; per Squid's SslPeekAndSplice wiki, without bumping, Squid sees the CONNECT destination and (after a peek) the SNI, and it logs no headers inside the tunnel. So the log from #37 through #39 would've recorded that curl reached a host, and it couldn't have shown what the User-Agent said.

That's a real hole in last week's argument, and I'd rather you hear it from me. A tripwire on the host fires on an off-list destination; it's got nothing to say about a good host receiving a bad header. A package registry is the kind of host that lands on an allowlist the first day, so the wire stays slack while a host you trusted receives your address. Nothing about the allowlist was wrong; the record underneath carries a blind spot the size of the request.

The transcript covers part of that blind spot. Anthropic's data-usage docs confirm transcripts are stored in plaintext under ~/.claude/projects/, which is where mt-empty's kind of finding comes from (those docs say nothing about the account email being given to the model or placed in headers). So the working practice is two records: the proxy log for where it went, the transcript for what the model wrote. Neither one answers the other's question.

The transcript has its own ceiling, and it's the one the commenter conceded. It catches only what the model wrote literally; an address computed at run time never appears as a string to find, and catching it takes interception or a listener on the far end. Interception in Squid means bumping TLS, and bumping needs a trusted CA on the client. That's the fair cost: a CA on the box, plus a brand-new log holding the headers you were worried about (congratulations, you've built a second thing to protect).

 

Let's Call It Manifest Engineering

A MANIFEST is the paper listing what's loaded on a truck; the guard at the gate logs the truck and the destination and never breaks the seal, so the paper is the only thing that says what's inside. For an agent, that's the written inventory of what the context holds about you and your org, kept next to the two records that show what left.

It's got three properties. It's written before loading: account, git config, CLAUDE.md, env, and whatever else the harness hands the model, each line checked against a plain test of whether you'd put it in an outbound header; the removal Anthropic announced is this step, done by the vendor on one field. It's reconciled against the gate: the proxy log says where the truck went, the transcript says what the model wrote on the way, and the two together have a shot at the good host receiving the bad header. And it's candid about cargo loaded after the paper's signed; a value the agent assembles at run time isn't on any manifest, so that case takes interception, on a test box, at the cost above.

There's a fair cost on the upstream side too, and Anthropic's own reason for adding the address is the proof. Take the value away and the model can go back to inventing names in @author tags. I'll take that trade; a fictional maintainer can be caught in review (and is the cheapest kind of maintainer you'll ever have), while an address in a request header is gone once it's gone.

Load only what you'd be willing to ship.

 

QUICK TIP

Grep Your Transcripts for Your Own Address

This pulls outbound-looking commands from Claude Code transcripts and keeps the ones carrying your address, masking the address in the output; run the same check with your git email if it differs from your account email.

ME='you@example.com'
grep -rEoh --include='*.jsonl' '(curl|wget|Invoke-WebRequest|urllib|requests\.(get|post)|fetch\().{0,200}' ~/.claude/projects \
  | grep -F -- "$ME" \
  | sed "s/$ME/<ME>/g" \
  | sort | uniq -c | sort -rn | head -20

It only finds what was written literally into a command, so a clean result doesn't clear an address built at run time.

 

Quick Wins

🟢 Easy (~10 min): Check your version against 2.1.317 and run the grep above. If you're on the stable channel, take note of the commenter's Nov. 6 line: "The stable channel is still on 2.1.311, so update to 2.1.317 or later."

🟡 Medium (~1 hour): List what your agent's context holds about you and your org (account, git config, CLAUDE.md, env), and remove anything you wouldn't put in an outbound header.

🔴 Advanced (half day): On a test box, trust a throwaway CA, bump TLS in Squid, log %{User-Agent}>h, and compare a day of headers against your identity list. Then tear it down, since the log it creates is sensitive.

 

Next Week

Next week, unless the news votes otherwise, tamper-evident records get the slot this issue bumped. The July agents from #39 tried to spoof or tamper with their own transcripts, and this week made the transcript one of the two records I'm asking you to lean on.

Keep out of the context what you wouldn't send.

Then read both records for what left.

The truck in this story carried one address, on an unproven charge of causing the scam mail, to one documented destination, and the paper trail was sitting in plaintext on our own disks the whole time; so write the manifest before the truck loads, keep the gate log and the transcript side by side, and treat every value you hand an agent as cargo that can leave the yard under a seal your guard doesn't open.

Bobby R. Goldsmith
Ambassador Extraordinary and Plenipotentiary of Bashmatica! by NodeBridge Automation Solutions

P.S. If someone forwarded this, subscribe at bashmatica.com (my list has your address too, and I promise to keep it out of my User-Agent).

NODEBRIDGE AUTOMATION SOLUTIONS

Every guardrail in this newsletter, already wired into your Claude Code setup.

NodeBridge configures Claude Code against your actual repos: MCP servers, subagents, hooks, permissions and a project memory that keeps the right context in and the stale context out. You own every config when it ships. Setups start at $750 and go live in about three days.

See the Setup Packages