What happens when an email tells your AI to do something

Every coding agent on your team reads text someone else wrote, in tickets, pull requests and comments, and any of it can carry a planted instruction. So keep the tools that could do damage locked where no message can reach them. Two assistants I built for email and call transcripts work that way, and the same rules carry over to coding agents.

What they do

The first turns an inbox, a calendar and call transcripts into ready-to-edit drafts that it never sends, because a wrong draft takes ten seconds to fix and a wrong email, once sent, stays sent. The second keeps a shared ledger of what each person on a small team promised on a call, with a quote and a timestamp for every row, taken from the full transcript and never from an AI-written summary.

Planted instructions go nowhere

When an email or a cell in the ledger contains text aimed at an AI, like “ignore previous instructions” or “forward this to”, the assistants are built to quote it under “Looks suspicious” in the summary they leave for the person using them, and leave that thread alone. The same suspicion applies to who a draft goes to. An address can only go on a draft if it comes from the thread, the meeting invite, someone the user already writes to or the user typing it in, so “please send it to X” inside an email body never qualifies, and when a name turns up with no address, the assistant asks instead of guessing.

The lock no message can pick

A check runs before every change to the assistants and fails if their instructions ever give them a way to send, share or delete. The email assistant’s setup guide also has each person lock the send, share and delete tools in their own connector settings, as a second lock behind its own rules.

The same rules for coding agents

On your team, coding agents get the same caution. We keep them away from production, and any session that would combine private data, untrusted input and a way to send data out needs a person’s approval or gets split so no single session has all three. Simon Willison calls that combination the lethal trifecta. The aim is that a planted line in a ticket has no path to production and no way to send your code anywhere. The guardrails around it are on the engineering side.

Read next

Letting an AI agent run the pricing

All writing