Email security checklist for teams with agents and humans

An email security checklist for a team with AI agents and humans covers six areas: authenticate your domains, limit what each agent and human can reach, screen mail before it is read, put a human on held mail and risky sends, check outgoing mail, and run drills. Start with authentication and agent access.

Updated

A team that runs agents on email faces two kinds of attack at once. One is the scam aimed at a human: the FBI IC3 2025 report counts $3.05 billion in business email compromise losses. The other is the instruction aimed at a model: OWASP ranks prompt injection as LLM01, the top risk in its Top 10 for LLM Applications, in both the 2025 edition and the current 2026 edition.

This checklist covers both readers. Each section says what to do. The links explain why.

1. Have you authenticated your domains?

Domain authentication stops others from sending mail in your exact name. It takes three DNS records for each domain. See SPF, DKIM and DMARC.

  • Publish one SPF record for every domain that sends mail, and keep it within the limit of 10 DNS lookups.
  • Sign all outgoing mail with DKIM.
  • Publish a DMARC record. Start at p=none and read the reports, then move to p=quarantine or p=reject.
  • Lock domains that never send mail with v=spf1 -all and a DMARC policy of p=reject.
  • Treat a pass on incoming mail as proof of the sending domain, not proof of safety. A lookalike domain passes.

2. Is access limited, for agents most of all?

Access is limited when each agent and each human can reach the mail they need and nothing more. An agent that is tricked can do no more than its access allows.

  • Give each agent its own inbox and address. Do not hand an agent a human's mailbox password or a shared mailbox.
  • Give each agent a key scoped to its job: one inbox, and read access before send access.
  • Keep login codes and password resets in a human's mailbox out of every agent's reach.
  • In a human's mailbox, let an agent write drafts and let the human send.
  • Turn on two-factor sign-in for every human, admins before anyone else.
  • Remove access the day someone leaves, and revoke keys that are no longer used.
  • Keep secrets and API keys out of email.

3. Is mail screened before it is read?

Mail should be screened before any reader sees it, because an agent reads within seconds and a cleanup scan after delivery comes too late. See pre-delivery vs post-delivery.

  • Screen before delivery for every inbox an agent reads.
  • Give an agent visible text alone: no hidden text and no raw HTML.
  • Hand mail to a model as data inside clear boundaries, never as instructions.
  • Hold doubtful mail outside the inbox, and make sure the agent cannot read the quarantine.
  • Screen attachments, including the text inside documents.

4. Is a human on the decisions that matter?

A human should make the calls an attacker most wants an agent, or a rushed colleague, to make alone. See human in the loop for email.

  • Name who reviews held mail, and how often.
  • Confirm any change of bank details by calling a number you already have.
  • Require a second human for payments above an amount you set.
  • Require approval before an agent's first message to a new recipient.
  • Make sure an agent cannot release held mail, loosen its own protection or allow a sender.

5. Is outgoing mail checked?

Outgoing mail needs checking because a hijacked agent does its damage by sending. A leak leaves through the same channel the attack came in on.

  • Screen what agents send for secrets and sensitive data.
  • Watch for replies steered to an address the inbox has never written to.
  • Limit how many recipients an agent can write to in a day.
  • Keep sent mail and blocked sends where a human can read them.

6. Do you run drills and keep records?

Drills show whether the controls above work before an attacker tests them. Records show what happened when one did not.

  • Send test attacks through the real mail path: a hidden instruction, a phishing link, an invoice fraud and a disguised attachment.
  • Run drills against agents and against humans.
  • Test again after any change to an agent's prompt, tools or access. OWASP's AI Agent Security Cheat Sheet recommends adversarial testing before deployment and after such changes.
  • Keep an audit log of releases, key changes and settings changes, and read it.
  • Write down what to do when something gets through: who to tell, which key to revoke, which sessions to end.

What changes between an agent and a human?

The same control looks different for each reader. An agent needs tighter limits, because it reads everything and may act on what it reads.

ControlFor an agentFor a human
IdentityIts own inbox and a scoped keyTheir own sign-in, with two-factor
ReadingVisible text alone, screened before deliveryScreened mail, with a warning on anything doubtful
Held mailSees that it exists. Cannot read or release it.Reviews it, then releases or discards
SendingScreened. A new recipient can wait for approval.Sends as usual. Confirms payment changes by phone.
Login codesIts own codes aloneKept out of every agent's reach

How much of this does Trotta Mail cover?

Trotta Mail builds many of the controls in sections 2 to 6 into the inbox, for agents and humans.

  • Every message is screened before delivery. An agent's key reads the safe view, and held mail waits for a human.
  • Text inside attached documents is screened like the message. An instruction aimed at an AI inside a document holds the message for an agent's inbox.
  • Keys are scoped. In a human's mailbox a key starts with drafts, and login mail starts sealed from every key.
  • An agent's key cannot release held mail, loosen its own protection or allow a sender.
  • Outgoing mail is screened, and the first message to a new recipient can be set to wait for approval.
  • Drills send realistic test mail through the real pipeline, and an audit log records who did what.

Some of the list stays with you: DNS records for domains hosted elsewhere, payment controls and offboarding. No filter catches everything, so the human steps are not optional.

The details are in the security model and on the security page. Every check is on every plan, and an inbox takes a minute to create.

Questions

What is the most important email security control?

There is no single one. For a domain it is DMARC at an enforcing policy, for humans it is two-factor sign-in and confirming payment changes by phone, and for agents it is screening mail before the agent reads it.

How is email security different for AI agents?

An agent reads every message within seconds and can treat text as an instruction. So mail has to be screened before the read, the agent should get visible text alone, and risky actions need a human.

Do small teams need DMARC?

Yes. Any domain that sends mail can be imitated, whatever its size. Gmail requires SPF or DKIM from every sender, and SPF, DKIM and DMARC from bulk senders.

Should an AI agent share a human's mailbox?

It is safer to give the agent its own inbox. If an agent must work in a human's mailbox, limit it to drafts and keep login codes and password resets out of its reach.

How often should a team run email security drills?

Run them on a regular schedule, and again after any change to an agent's prompt, tools or access. A drill is cheap, and a control that was never tested is a guess.

Sources

Give your agent an inbox of its own. Free for two inboxes. No card.