Skip to content

Helper agents put my email in more than 150 web requests

Nobody asked them to, and Claude Code's own instructions said not to. What my logs showed about leaked email, keys every agent can read and rules that reach only one agent, and what stopped it in 294 tests.

Matt WaiteUpdated 7 min read

Between Oct 22 and Nov 6, helper agents on my Mac put my email address into more than 150 web requests. They went mostly to museums and image archives: Wikimedia, the Met, the Art Institute of Chicago, the Rijksmuseum and the Smithsonian. Nine organizations got it in all. Nobody asked the agents to do it. Claude Code's own instructions said not to.

I never saw any of it. All 58 commands that sent it came from helper agents: Claude Code's subagents, which it started in the background to split up a job.

The harm was small. Some museums and archives have my address in their server logs. Nothing was bought or signed up for. But I found out two weeks later, by accident, and the day after I found out, more than 50 more requests went out.

A written instruction didn't stop it. In my tests, what did stop it was a check at the point of action.

What Claude Code told them

Claude Code gave each session the signed-in user's email, with an instruction. On my Mac it read:

The user's email address is [my email]. Use it only to identify the user, such as for authorship, attribution, or filtering their own work. Never send it to an unrelated service, such as in a request header, URL, or payload, unless the user explicitly asks.

Every helper that sent my email had that instruction. In the first one, my address went out 59 seconds after the instruction arrived, in a header the agent named itself: ExcellentBrandStudy/1.0 ([my email]).

Only one helper's instructions mentioned my email at all, and only as the author for its git commits. None of my rules or notes asked for it either way.

Why, probably

These sites ask scripts to say who they are. Wikimedia's User-Agent policy asks for contact details "as an email address, a website, or a wiki user", and says scripts without them "may be blocked without notice". The helpers followed the website's etiquette over Claude Code's instruction. They answered, politely, with my address.

I can't say that for certain. It's the reason that fits every case.

Anthropic has changed it since. From Nov 11, about 19 in 20 new sessions on my Mac no longer get my email, and the rest get firmer wording. The instruction was a good one. It just wasn't enough on its own.

When it happened

Most of it came from one afternoon. On Oct 22, helper agents working on a brand study for Excellent looked for reference images in museum archives. 9 of the study's 30 helpers sent my email: 30 of the 58 commands. Five more commands went out on Nov 2, from a research helper. The last 23 came on Nov 6, from four helpers: one downloading organizations' logos, one making brand prints, one finding a photo for my portfolio site, and one reviewing that site.

I found out on Nov 5.

What else the logs showed

The email is the small version. Here's the rest, Oct 21 to Nov 18:

  • Most of the work ran in helpers I never watched. Helper agents wrote 796 of the 1,106 session logs Claude Code kept.
  • A browser agent used my signed-in Chrome. It took more than 2,400 actions there, on 118 sites, including my email inbox. It typed my email into a website 8 times. Most of these were jobs I asked for.
  • My rules didn't reach every agent. 15 rules I wrote reach only one of my two main coding agents, Claude Code and Codex. One of them, my list of folders that must never be deleted, reaches Claude Code but not Codex. I wrote it after a cleanup deleted folders I needed.
  • Every agent can read my keys. 53 keys on this Mac are readable by every agent. None was found sent to a model provider.

None of this needed an agent to go rogue, or me to be careless. Agents now move faster than anyone can manage them: hundreds of steps, many in helpers, while you're doing something else.

What stopped it in tests

I wanted to know what actually holds. I ran 294 live tests with real Claude Code and Codex agents, each in a fake home with a fake owner. Rules the agent never saw were broken in 14 of 24 runs. When the rule did reach the agent, it was followed almost every time.

Then I moved the rules to where things happen. The operating system refused the delete. The outgoing request was checked before it left. The git server refused to rewrite its history. With all of that in place, no rule was broken in 72 of 72 runs, including 48 where the agent was told to get around the rules.

The full method and tables are in What stops an agent from breaking a rule.

What I think should be true

  1. You can see what your agents did. Every agent and every helper, in plain words, without reading a log.
  2. Your rules hold where the action happens. The ones that matter shouldn't depend on an agent remembering them.
  3. Your logs don't have to leave your machine to do it. A tool that watches your agents shouldn't become one more place your email and keys end up.

None of this is a case against the people who make agents. I use them all day and I'll keep using them. But they do real work on our machines now, and anything that does needs two things: a record you can read, and a lock that doesn't depend on good behavior.

So I built Excellent

Excellent for Mac, for Macs with Apple silicon and macOS 14 or later, reads the logs your agents already keep, on your Mac, and shows what went wrong: your email sent to websites or typed into forms, keys sitting in logs, secrets files read, history overwritten, and rules that reach only one agent. Seeing all of that is free: scan any time and see every finding.

It can also share your rules between Claude Code and Codex: it shows which reach only one of them and, on your OK, adds them to a file the other reads.

Excellent Pro checks risky actions before they run, in 9 agents. It blocks your email on its way to a website and an API key typed into a command, and asks before a force push, skipping your git hooks or a look at your keys. It can turn your file and folder rules into the Mac's own sandbox for Claude Code and Codex, which holds however a command is worded. It scans in the background and alerts you. Pro is $6 a month or $60 a year in the US. See pricing and current availability.

Of the three places my tests enforced rules, Excellent for Mac can set up all three: the Mac's own sandbox and a check on what agents' commands send to websites, both with Pro, and branch protection on GitHub or GitLab. That check is off until you turn it on in Settings. It covers commands that use its proxy settings, not tools like gh, Homebrew or terraform. That's Excellent for Mac today; a version for Linux servers comes next, then Windows.

The free app needs no account; Pro needs a sign-in. Unless you sign in for Pro, share usage stats or pair it with our team, nothing about you or your agents is sent to us. How Excellent for Mac works, and what it sends has the details and what it can't do.

Check your own Mac

  • Get the app. Excellent for Mac is free to scan with. A scan takes seconds.
  • Or ask your agent. The free scan is one prompt for Claude Code or Codex. It runs a small read-only script on your logs and writes one report on your Mac.
  • Or do it by hand. Check what your AI agent actually did has the searches, including the one for your email in web requests.

The AI agent security guide shows how to check your own sessions and choose a control for what you find.

How I counted

  • What was read. The session logs AI agent tools keep on my Mac, read where they sit by Excellent's own engine, with its results kept in a scratch folder. Nothing in the logs was changed. In all, 5,945 logs and 161,420 tool calls from three agent tools: Claude Code, Codex and Grok Build. Claude Code's logs only went back about four weeks, because it deletes them after 30 days by default. I took a snapshot of the logs on Nov 18 so the numbers can be re-run.
  • What a count is. A command an agent ran that carried my email to a web address, each counted once. Some helper logs start as copies of another helper's history, so the same command can sit in several logs. A request counts when the log shows the server's answer, error answers included. Later runs of scripts that held my email don't show it in the command, so they aren't in the 58. With them, it was more than 200 commands.
  • What "nobody asked" means. I read the instructions of all 14 helpers that sent my email. Four were copies of a fifth and share its instructions. Only one mentioned my email, as its git commit author. No instruction file or memory note on the Mac mentions it either way.
  • What it can't say. It's one Mac, a heavy user's: a case study, not a rate. My helpers did it unasked, in long sessions with many helpers at once. In my first short test, ordinary agents didn't send the email (0 of 6 runs). In a second, when Claude Code and Codex got the email with the instruction quoted above, neither sent it (0 of 14). Short tests don't model those long sessions. "Can read" isn't "leaked": no key was found sent to a model provider. The browser agent's site categories are rough.

Install Excellent