Glossary
The terms behind email security and AI agents, in plain words.
A
Account takeover
Account takeover is when an attacker gains control of a real account, usually with a stolen password, a stolen session or an intercepted login code. Mail sent from a taken-over mailbox passes authentication checks, because it really does come from that account.
Agent hijacking
Agent hijacking is an attack that turns an AI agent against its owner. The attacker plants instructions in content the agent reads, such as an email, and the agent then uses its own access and tools to do what the attacker wants.
Agent inbox
An agent inbox is an email inbox made for an AI agent to use through an API instead of a mail app. The agent has its own address, receives and sends mail in code, and can be limited in what it reads and where it sends.
AI agent
An AI agent is software built on a large language model that takes actions, not just produces text. It reads input, decides what to do and uses tools such as email, search or code to do it, often with no human checking each step.
Allow list
An allow list is a set of senders or recipients that a mail system treats as trusted. Mail from an allow-listed address skips some checks or passes a strict mode. Because it lowers protection, adding to an allow list should be a human decision.
Attachment sandboxing
Attachment sandboxing is opening an email attachment in an isolated environment before delivery to watch what it does. If the file tries to install software, change settings or contact a remote server, it is treated as malicious and kept from the recipient.
B
Business email compromise (BEC)
Business email compromise (BEC) is fraud carried out by email that impersonates someone trusted, such as an executive, a supplier or a lawyer, to get money or data sent to the attacker. It often uses no malware and no link, just a believable request.
C
CEO fraud
CEO fraud is a form of business email compromise in which the attacker poses as a senior executive. The message asks an employee, often in finance, for an urgent and confidential payment or for gift cards, and leans on authority to skip the normal checks.
Credential phishing
Credential phishing is an attack that tricks a recipient into entering a username and password on a fake sign-in page. The email usually imitates a known service and links to a copy of its login screen, which sends whatever is typed to the attacker.
D
Data exfiltration
Data exfiltration is the unauthorized movement of data out of a system. With AI agents, it often means an agent being steered into sending private information to an attacker, through an email, a link it fetches or an image address that carries the data.
DKIM
DKIM (DomainKeys Identified Mail) is an email authentication standard that adds a digital signature to a message. The receiver checks it against a public key in the signing domain's DNS. A valid signature shows that the domain signed the message and that the signed parts were not changed.
DMARC
DMARC (Domain-based Message Authentication, Reporting and Conformance) is a standard that ties SPF and DKIM to the address in the From line. A domain publishes a policy in DNS that tells receivers what to do with mail that fails, and where to send reports.
DNS
DNS (the Domain Name System) is the directory of the internet. It maps domain names to the records computers need, such as server addresses. Email depends on it: MX records say where a domain's mail goes, and TXT records hold SPF, DKIM and DMARC.
E
Email authentication
Email authentication is the set of standards a receiving server uses to check that a message really comes from the domain it claims. The three in common use are SPF, DKIM and DMARC. Authentication proves where mail came from, not that it is safe.
Email spoofing
Email spoofing is forging the sender of an email so it appears to come from someone else. The attacker fakes the From address or the display name. SPF, DKIM and DMARC make exact-domain spoofing detectable, but not lookalike domains or forged display names.
H
Held mail
Held mail is email that was accepted but kept out of the inbox until someone decides whether to release it. In Trotta Mail, held mail waits for a human, and an agent's API key sees that the message exists and why, with nothing its sender wrote.
Hidden text
Hidden text is content in an email that a human reader cannot see but software can read, such as white text on a white background or text set to zero size. Attackers use it to pass instructions to an AI agent without the human noticing.
Human in the loop
Human in the loop means a human must review or approve a step before an automated system carries it out. For an AI agent on email, the usual points are releasing held mail and approving a message to a new recipient.
I
Indirect prompt injection
Indirect prompt injection is a prompt injection that arrives inside content an AI model reads, not from whoever gave the model its task. The instructions are planted in an email, a web page or a file, and the model treats them as commands.
Invoice fraud
Invoice fraud is a scam in which an attacker sends a fake invoice, or changes the bank details on a real one, so that a payment goes to the attacker's account. It often impersonates a known supplier and arrives when a real invoice is expected.
L
Large language model (LLM)
A large language model (LLM) is an AI system trained on large amounts of text to predict and produce language. It powers chat assistants and AI agents. An LLM cannot reliably tell instructions from data in the text it reads, which is what makes prompt injection possible.
Lookalike domain
A lookalike domain is a domain name registered to resemble a real one, by swapping, adding or dropping a character or changing the ending. Mail from it can pass SPF, DKIM and DMARC, because the attacker owns the domain and can publish valid records.
M
MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard for connecting AI applications to outside tools and data. An MCP server offers tools, such as reading or sending email, and an MCP client, such as a chat assistant or a coding agent, calls them.
MX record
An MX record is the DNS record that says which servers accept email for a domain. A sending server looks up the recipient domain's MX records and tries the one with the lowest preference number before the others. Changing a domain's MX record changes where its mail is delivered.
O
OAuth
OAuth is a standard that lets one application act for an account without learning its password. The account owner signs in at the service and approves specific permissions. The application receives a token limited to those permissions, which can be revoked later.
One-time code
A one-time code is a short code, usually a few digits, sent by email or text to confirm a sign-in or a sign-up. It works once and expires quickly. Whoever can read the message can use the code, so it needs protecting like a password.
Outbound screening
Outbound screening is checking a message before it leaves, not just when one arrives. It looks for secrets, sensitive data and signs that a sender, human or AI agent, is being steered into sending information to the wrong recipient. A risky message is blocked or waits for approval.
P
Payroll diversion
Payroll diversion is a scam in which an attacker poses as an employee and asks payroll or human resources to change the bank account their salary is paid into. The next pay run then goes to an account the attacker controls.
Phishing
Phishing is a message that pretends to come from a trusted source to trick the recipient into giving up something of value: a password, a payment, data or access. It arrives by email, text message, phone call or QR code.
Post-delivery remediation
Post-delivery remediation is finding and removing a harmful email after it has reached a mailbox. A security tool rescans delivered mail as new information arrives and moves or deletes what it now knows is bad. It cannot undo a read, a click or a reply.
Pre-delivery screening
Pre-delivery screening is checking an email before it reaches an inbox, while the sending server is still connected. Depending on the result, the message is refused, held outside the inbox for a human, or delivered. Nothing unscreened is readable at any moment.
Prompt injection
Prompt injection is an attack that feeds an AI model text written to override its instructions. The model follows the attacker's words instead of its owner's. OWASP lists it as LLM01, the top risk in its Top 10 for LLM Applications, in both the 2025 and 2026 editions.
Q
Quarantine
Quarantine is a holding area outside the inbox for email that a filter judged dangerous or unwanted. The message was accepted but not delivered. Someone with the right permission reviews it, then releases it to the inbox or deletes it.
Quishing
Quishing is phishing that uses a QR code in place of a link. The code sits in an email or an attachment, and scanning it with a phone opens a fake sign-in or payment page, often on a device outside the company's protection.
S
Safe view
The safe view is the version of a message that Trotta Mail gives an agent's API key. It contains what a human could see and nothing else: hidden text is removed, instructions aimed at a model are replaced by a note, and links that are not clean are defanged.
Secure email gateway
A secure email gateway (SEG) is a filter that sits in front of a mail server and screens messages before passing them on. A domain's MX record points to the gateway, which blocks, quarantines or relays each message. Many gateways also check outgoing mail.
SMTP
SMTP (Simple Mail Transfer Protocol) is the standard servers use to pass email to each other. The sending server connects, names the sender and the recipient, and transmits the message. The receiving server answers each step with a numeric code that accepts, defers or refuses.
Spear phishing
Spear phishing is phishing aimed at one chosen target. The attacker researches the recipient and writes a message that refers to real names, projects or events, so it reads like ordinary business mail and is harder to spot than a mass campaign.
SPF
SPF (Sender Policy Framework) is an email authentication standard that lets a domain list, in a DNS record, the servers allowed to send its mail. The receiver compares the connecting server with that list. SPF checks the envelope sender, not the From line a reader sees.
T
Thread hijacking
Thread hijacking is replying inside a real email conversation from an account or a copy the attacker controls. Because the message quotes genuine earlier mail and arrives in a trusted thread, a harmful link, file or payment request in it looks routine.
V
Vendor email compromise
Vendor email compromise is business email compromise that starts with a supplier. The attacker takes over or imitates a vendor's mailbox, studies its billing, then sends that vendor's customers real-looking invoices or new bank details.
Verdict
A verdict is the decision a mail filter records about a message. In Trotta Mail every message, in both directions, carries a verdict made before the message is readable: an action (deliver, flag, hold or reject), a level, and the reasons in plain words.
W
Webhook
A webhook is an HTTP request that a service sends to an address you choose when something happens, such as a new email arriving. It replaces polling: your code is called when an event happens, and a retry can deliver the same event again. Webhook requests should be signed so the receiver can verify them.
Z
Zero-click attack
A zero-click attack is one that needs no action from its target: no link clicked and no file opened. With AI agents, receiving the message can be enough, because the agent reads it and acts on hidden instructions by itself.