What is pre-delivery email screening?
Pre-delivery email screening is checking a message before it can be read, while the sending server is still connected. The receiving system inspects the sender, the links, the files and the text during the SMTP conversation, then refuses the message, holds it outside the inbox for a human, or delivers it.
Updated
What does pre-delivery mean?
Pre-delivery means the decision about a message is made before the message reaches an inbox. Nothing unscreened is readable at any moment, by a human or by an AI agent.
The opposite order is post-delivery: the message lands, and a later scan removes it if it turns out to be bad. Pre-delivery vs post-delivery covers what each order can and cannot do.
Pre-delivery screening is not new. Gmail and Outlook filter spam and malware before delivery, and secure email gateways sit in front of mail servers to do the same. Google Workspace has a setting named enhanced pre-delivery message scanning, which slightly delays a suspicious message for extra checks.
The differences between systems are in when the decision is made, what is checked, and where a doubtful message waits.
What happens during the SMTP conversation?
Email moves between servers over SMTP, a short scripted conversation defined in RFC 5321. The sending server connects, names the sender and the recipient, sends the message, and waits for one final reply.
S: 220 mx.example.net ESMTP
C: EHLO mail.northwind-supply.example
S: 250 mx.example.net
C: MAIL FROM:<billing@northwind-supply.example>
S: 250 OK
C: RCPT TO:<riley@example.net>
S: 250 OK
C: DATA
S: 354 Start mail input; end with <CRLF>.<CRLF>
C: (headers and body)
C: .
S: 250 OK accepted
or
S: 550 5.7.1 Message refused refused
C: QUIT
S: 221 ByeThat last reply is the handoff. RFC 5321 says a receiver that answers 250 after the message data accepts responsibility for delivering it. A reply that starts with 5 is a permanent refusal: the receiver takes nothing, and the sending server reports the failure to its own sender.
Pre-delivery screening in the strict sense happens between the final period and that reply. The receiver has the whole message, and the sender is still on the line.
| Point in the conversation | What the receiver has | What it can check |
|---|---|---|
| Connection | The sending server's address | Reputation and connection limits |
MAIL FROM and RCPT TO | The envelope sender and the recipient | SPF, block lists, whether the recipient exists |
After DATA | The whole message | DKIM, DMARC, the visible sender, links, attachments and the text |
What are the three outcomes?
Pre-delivery screening ends in one of three outcomes: refuse, hold or deliver.
| Outcome | What the sending server is told | Where the message goes | When it fits |
|---|---|---|---|
| Refuse | A permanent error, such as 550 5.7.1 | Nowhere the recipient can read it | The message is malicious |
| Hold | Accepted, 250 | A quarantine outside the inbox, until a human releases it | The message looks dangerous and a human should decide |
| Deliver | Accepted, 250 | The inbox, often with a verdict or a warning attached | Nothing was found, or the finding is minor |
Accepted does not mean delivered. A receiver can answer 250 and still keep a message out of the inbox. That is what a hold is, and held mail explains who may release it.
Refusing is more honest than deleting in silence. The sending server learns at once, so a legitimate sender finds out and can try another way. RFC 5321 warns that silently dropping mail after accepting it is dangerous, because senders expect mail to be delivered or returned.
What does pre-delivery screening check?
Pre-delivery screening checks who sent the message, what the message contains, and who the text is written for.
- The sender. SPF, DKIM and DMARC results, lookalike domains, and whether this sender has written before. See SPF, DKIM and DMARC.
- The links. Where a link really goes, and whether its text claims something else.
- The files. What an attachment really is, whatever its name says.
- The request. Signs of payment fraud, or a page that asks for a password.
- Text aimed at a model. Hidden text and instructions written for an AI reader, which matter when an agent reads the inbox.
Why does timing matter more for AI agents?
Timing matters more for AI agents because an agent reads a new message within seconds and may act on what it says, so a later scan has no window to work in. Pre-delivery vs post-delivery walks through both orders.
An agent also takes text literally in a way a human does not. A line such as "ignore your previous instructions" is an attack on a model. Screening for an agent has to look for prompt injection, and it has to finish before the agent's first read.
The volume is real. Phishing and spoofing was the most reported crime type in the FBI IC3 2025 report, with 191,561 complaints.
What are the limits of pre-delivery screening?
No filter catches everything, and pre-delivery screening is no exception. It judges a message once, with what is known at that moment.
- A link can point to a clean page at delivery and a harmful one an hour later.
- A scam written in plain, polite words with no technical tell can pass.
- Mail from a real account that was taken over authenticates correctly.
- The check must be quick. The sending server is waiting, and RFC 5321 asks receivers to keep the wait for the final reply short.
This is why uncertain mail should be held for a human and not guessed at. It is also why a cleanup scan after delivery is still worth having as a second layer for human mailboxes.
How does Trotta Mail do pre-delivery screening?
Trotta Mail is an inbox that screens at its own door. Every message is screened while the sending server is still connected, for an agent's inbox or a human's, including mail between Trotta Mail inboxes.
- Malicious mail, known or new, is refused with an error. A copy waits in Blocked for a human, in case it was a mistake.
- Dangerous mail is accepted and held outside the inbox until a human releases it. An agent's API key cannot read it.
- Mail worth a second look is delivered with the caution attached.
- Everything else is delivered with a verdict that says nothing was found.
A refusal does not tell the sender why, so an attacker is not told what gave the message away. The reason stays with the refused message, for the inbox owner.
Each message carries a verdict with the SPF, DKIM and DMARC results and the reasons in plain words. The guard is stricter for an agent than for a human, and outgoing mail is screened too. The rules are in concepts and the security model.
Every check is on every plan, including the free one. Create an inbox to see a verdict on your own mail.
Questions
Is pre-delivery screening the same as a spam filter?
A spam filter is one kind of pre-delivery screening. Security screening goes further: it looks for fraud, harmful links and files, and text aimed at an AI agent, and it can refuse or hold a message instead of filing it in a folder.
Does the sender know when a message is refused?
Yes. The sending server gets a permanent error during the SMTP conversation and reports the failure to whoever sent the message. A held message is different: the sending server is told it was accepted.
Do Gmail and Outlook screen mail before delivery?
Yes. Both filter spam and malware before a message reaches the inbox. A message that looks real can still pass, and neither was built for an AI agent reading the mail.
Does pre-delivery screening slow email down?
Slightly. The check runs while the sending server waits for the final reply, so it has to finish quickly.
Can pre-delivery screening stop prompt injection?
It can stop a great deal of it when the screen looks for hidden text and instructions aimed at a model. No filter catches everything, so an agent should still treat mail as data, never as instructions.