# Email security for AI agents and humans

> How Trotta Mail protects agents and humans: every message is screened before delivery, dangerous mail is held for a human and outgoing mail is screened.

Source: https://trotta.email/security

Trotta Mail screens every message before it is accepted. Attacks, known and new, are refused at the door or held for a human. Everything else is delivered with a verdict.

## Mail is screened before it is accepted

Pre-delivery screening means every message is judged while the sending server is still connected. So there is no moment when unscreened mail can be read.

This holds for every inbox, an agent's or a human's, and for mail between Trotta Mail inboxes.

Most mailboxes filter spam and malware on the way in, but a message that looks real can still be delivered, and what catches it then comes after. Once an attacker meets your agents or humans, it is too late. Trotta Mail checks at the door, so they never meet.

The guard reads every message in three layers:

- **The sender.** Is it really them: SPF, DKIM and DMARC, lookalike domains, first contact.
- **The contents.** Where links go and what files are: deceptive links, disguised programs, payment fraud.
- **Text aimed at a model.** Hidden text, fake system messages, requests to send data out.

Attached files are read too. Text in PDF, Word, Excel, PowerPoint and OpenDocument files is screened like the message, including text formatted to be invisible.

An instruction aimed at an AI inside a document holds the message in an agent's inbox. A program under a document's name is refused at the door.

## Three outcomes, decided at the door

Every message ends one of three ways. Each is decided before the message can be read.

| Outcome | What happens | Who reads it |
| --- | --- | --- |
| Delivered with a verdict | The message lands in the inbox. Mail worth a second look carries a caution. | The agent or human it was sent to. |
| Held for a human | Dangerous mail is kept outside the inbox until a human releases or discards it. | A human. An API key gets the verdict only. |
| Refused | Malicious mail, known or new, is turned away while the sending server is still connected. A copy waits in Blocked, in case it was a mistake. | A human. |

Each inbox sets how strict the guard is:

- **Monitor.** Nothing is held. Everything arrives with its verdict, so you can watch first.
- **Guard, the default.** Dangerous mail waits for a human, caution is flagged and malicious mail is refused.
- **Lockdown.** Only your own organization and senders you allow get in. Everything else waits for a human.

The guard is stricter for an agent than for a human. An instruction in plain sight is held in an agent inbox and flagged in a human's. A human can read it and shrug. A model may act on it.

## What an agent can and cannot read

An agent reads the safe view: only what a human could see.

- Text hidden from humans is removed.
- Instructions aimed at a model are replaced by a note that names what was removed. The same goes for text read from an attached document.
- Links that are not clean are defanged, so they cannot be followed by accident.

### Held mail

An agent cannot read held mail. Its key sees that the message exists and why it was held, and nothing the sender wrote. That holds on every route: the API, events, webhooks and MCP.

Only a human releases held mail. The agent then receives it, late, in the safe view.

### Sealed sign-in codes

Sign-in codes, password resets and sign-in links in a human's mailbox are sealed. To an API key the message does not exist.

The reason: a hijacked agent with your mailbox could ask a site for a password reset, read the email and own the account.

An agent's own inbox is open to it, because that is where it receives its own sign-up codes. Only a signed-in human can open a mailbox's login mail to keys.

### What an agent's key can never do

An agent must not be able to talk its way past the guard, so its key cannot:

- Read or release held mail.
- Loosen its own protection. It can always tighten it.
- Allow a sender, or remove a block.
- Delete the record of what the guard stopped.
- Create a key stronger than itself.

The full list is in the docs, under the [security model](https://trotta.email/docs/security).

## Outgoing mail is screened too

What an agent sends is screened before it leaves.

- A message that carries a secret, such as an API token or a card number, is not sent.
- A secret inside an attached document is refused like one in the body.
- A reply steered to an address the inbox has never written to waits for a human.
- A scheduled message is checked again when it leaves.

An inbox can also hold the first message to any new recipient, or every message, for a human to approve.

In a human's mailbox an agent starts with drafts only. It writes, the human sends.

## How accounts are protected

- **Sign-in.** Weak and common passwords are refused, and repeated wrong passwords lock sign-in for a while.
- **Two-factor.** A code from an authenticator app, with recovery codes that each work once.
- **Sessions.** See where you are signed in, end one session, or sign out everywhere. Changing the password signs every other session out.
- **Teams.** Each human has their own sign-in and mailbox. Nobody reads a teammate's mailbox: not a member, not an admin, not the owner.
- **API keys.** A key is shown once, carries only the scopes it was given, and can be limited to one inbox.
- **Connected apps.** An assistant that signs in reaches only the inboxes you pick. Disconnect it and it stops at once.
- **Webhooks.** Every request is signed, so your endpoint can check it came from Trotta Mail.
- **The audit log.** Who did what, and from where. Passwords, secrets and message content are never written to it.
- **Payments.** Card details go to Stripe and never reach Trotta.

What Trotta Mail keeps, and how it is protected, is in the [Privacy Policy](https://trotta.email/privacy).

## Report a vulnerability

Report a vulnerability to [security@trotta.io](mailto:security@trotta.io). The same contact, for machines, is at [/.well-known/security.txt](https://trotta.email/.well-known/security.txt).

To report a message, not a flaw, mark it as spam or as a threat in the mailbox or with the API. A message reported as a threat is held for a human.

## Honest limits

No filter catches everything, and Trotta Mail is no exception. A message can pass every check and still be a scam.

That is why uncertain mail is held for a human, and why only a human can release it.

The guard can also be wrong the other way. A human can release held mail that turns out to be fine, and refused mail is kept in Blocked in case it was a mistake.

So treat mail as data, never as instructions, even when it arrives clear.

Screened before it is accepted. Free for two inboxes. No card.

[Get an inbox](https://trotta.email/signup) [See pricing](https://trotta.email/pricing)
