Email consists of three planes: the attack plane, the data plane, and the control plane. Learn how to secure them all.
Conversations about safely enabling agentic access to inboxes usually focus on prompt injection. If an agent is processing every email coming in, then attackers get an easy way to deliver malicious instructions that the agent will happily carry out. But prompt injection only covers the attack plane. Within email, there are two additional planes of risk to consider: the data plane and the control plane. In other words, what data can an agent access within an inbox and what tasks does that access make possible?

Securely connecting agents to email requires covering all three planes. The good news is that an agent inside a mailbox can be modeled as an adversary inside a mailbox, and the same controls we built to contain one apply to the other.
Why agents want access to email
From personal assistants like Muse, Instinct and Town to enterprise apps like Claude, ChatGPT and Gemini, every AI app wants access to email. The reason is simple: email holds a massive amount of context on an individual or business. The mailbox captures what’s top of mind and becomes a system of record for what was decided, planned, and actioned, and with whom. Even when the context lives elsewhere, it finds its way to email via notifications and digests. An agent without mailbox access, therefore, is fundamentally limited in its ability to be helpful.
So individuals and businesses are going to connect agents to email, and most already have. The question is how to enable that safely.
The three planes of risk
The attack plane is the most commonly cited, and for good reason. With AI agents in the mix, attackers have gained a new type of payload that is both more powerful and harder to detect. Messages carry hidden instructions that users never see, and agents read them and perform actions the user never authorized. This isn't hypothetical. In 2025, Aim Security disclosed EchoLeak, a flaw in Microsoft 365 Copilot that let an attacker pull sensitive data out of a victim's Microsoft 365 environment simply by sending a single crafted email.

But email is not just a way of exchanging messages, it’s also a content repository, and you also have to threat model the data plane. There is no way to grant an agent part of a mailbox: the access model in Gmail or Outlook is all or nothing. That’s a problem given that mailboxes tend to hold years of sensitive information like legal and financial documents, personnel data, and internal memos. Agents can run search queries, access these emails on demand, and store them elsewhere as they see fit. It’s no different than any other file repository. Attackers have known what a mailbox is worth for years and losing the data within them can be very costly. An attack on PIH Health a few years ago compromised 45 employee mailboxes holding the protected health information of nearly 190,000 patients, and ended in a $600,000 settlement with HHS's Office for Civil Rights.
Lastly, most people forget that email is also a control plane. The ability to access email is a proxy for identity verification. Password resets, magic sign-in links, and account verification emails all assume that whoever can read the inbox is you. Connect an agent to that inbox, and you have transferred that power to it. Again, this attack path is well trodden. Earlier this year, an attacker used a compromised Gmail OAuth token belonging to a Composio employee to intercept magic-link sign-in emails, and from there reached production systems holding customer secrets.
In each case, attackers have exploited the multifaceted nature of email in the past. There is no reason to believe malicious or compromised agents won’t do the same.
Detect the attack, gate the access
The attack plane is a detection problem and the traditional email security model can be carried over: inspect incoming emails, detect malicious instructions, and remove messages before they are processed by an agent. But the data and control planes need a different kind of control. When an agent reads a sensitive thread or follows a password reset link, there is nothing malicious to detect. The answer there is gating: granular control over what an agent can see and do, with a human in the loop for anything sensitive. Granular access control is also the backstop for the attack plane. If an injected instruction slips past detection, the steered agent still can't reach the protected data or complete the reset.
At Material, we go beyond traditional email security and have always thought about email from all three lenses. The controls we have built over the years to stop attackers are exactly the same controls necessary to limit agents:
- Attack plane: Our updated detection engine uses both heuristics and models to find prompt injection techniques in incoming mail and remediate matching messages.
- Data plane: We scan the entire mailbox and redact sensitive messages at rest, so an agent can't read them without a human-in-the-loop escalation.
- Control plane: We intercept password resets and other identity-related emails and force an out-of-band MFA challenge before the message becomes visible, effectively putting a human in the loop for every identity-related transaction happening over email.

Secure the resource, not the actor
The deeper lesson reaches beyond email. Strong security controls are effective whether the actor is an attacker or an agent, because they key on what is being touched rather than who is touching it. An authorized agent and an attacker holding a stolen token look the same to a mailbox, and the same is true of a file share, a calendar, or anything else an agent can reach.

As agents spread across the enterprise, the teams that already know how to contain a compromised account will find they've been building agent security all along.
Ready to securely connect agents to email? Get a demo of Material today.

