The email that phishes you and prompts your AI

October 10, 2026 · SPR{K3 Research

On October 7, Barracuda's threat research team published a Threat Spotlight that is worth reading closely if your organization has stood up an AI email assistant in the last year. The analyst, Guruprasad Kenja, describes a current campaign in which a single phishing email is doing two jobs at once: delivering a password-protected attachment aimed at the human recipient, and delivering a hidden instruction payload aimed at that recipient's AI assistant. One envelope, two targets, two different control planes.

The reason this matters is not that prompt injection in email is new. Microsoft 365 Copilot's EchoLeak chain and Salesforce Agentforce's SalesBleed chain already made the case that email-shaped inputs are reaching LLM-shaped outputs without passing through anything that treats the hidden text as instructions. What Barracuda's writeup adds is a named, enumerated primitive list, published by a tier-1 email security vendor, that puts specific numbers on the shape of the attack surface.

Four hiding techniques, one gap

Barracuda lists four techniques the campaign is using to hide instructions from the human reader while leaving them visible to the AI assistant:

None of these are hard to detect in isolation. The problem is that the control plane doing the detecting at a mail-security-gateway tier was built to answer a different question. It was built to decide whether the attachment is malware, whether the sender's domain is spoofed, whether the message is spam. It was not built to ask whether some invisible text in the body is being read by a downstream agent as instructions. The AI assistant, meanwhile, treats the entire message body as content and does not have a notion of "this part was meant to be visible to a human and that part wasn't."

That gap is where the campaign lives.

What the AI is being told to do

Barracuda also lists five specific downstream behaviors the hidden instructions are driving. Each one is worth sitting with for a moment, because they aren't the stereotypical "ignore previous instructions" examples:

The first three are business-email-compromise variants. BEC is a problem enterprises have funded mitigations against for years, and the mitigations work when the human reader is the thing the attacker needs to flip. The AI assistant is a different thing to flip, and the mitigations largely don't follow.

The last two — the chatbot primitive and the poisoned-documentation primitive — are more structural. They are the public, named-vendor-published articulation of two patterns that have been building through 2026: that any public-facing chatbot's own confirmation and configuration UI is itself an attack surface, and that the writable text surfaces an AI coding agent reads at job time are code-bearing. Pillar Security's Deadbugz campaign made the same case against MCP server metadata. HackerNoon's Install-Guide piece made it against README files. Barracuda's Threat Spotlight is now making it against email attachments and inline documentation.

Why this is a boundary problem, not a filter problem

It is tempting to read "AI assistant ingests hidden text" and reach for a filter. Strip HTML comments. Decode Base64. Normalize zero-width characters. Those are all defensible controls at the content-rendering layer, and Barracuda explicitly recommends them: input sanitization, prompt-injection detection, treating external content as data only and separating it from instructions.

But the deeper observation in the piece is architectural. The twenty-year trust boundary that email security has policed — what the human reader sees, what the sandbox sees, what the attachment scanner sees — is not the same boundary the AI assistant is policing against. The AI reads a superset of the message. It reads the parts a human cannot see. It takes the whole thing as its input context and acts.

The practical consequence is that two different control planes are now making security-relevant decisions about the same piece of attacker-controlled text, with no shared view of what was taken as data versus what was taken as instructions. The mail security control plane sees "an email with an attachment, scanned, verdict delivered." The AI assistant control plane sees "a prompt, including hidden text, acted on." The attacker can shape the message so that both verdicts are the one they want.

Barracuda's own framing lands on the same point: "treat external content as data only, separate from instructions." The sentiment is right. The architectural problem is that the AI assistant's internal boundary between "task context from the user" and "message content from the outside" is porous — the model fundamentally reads them as the same substrate — and no post-hoc filter at the gateway closes that gap unless it has visibility into what the AI subsequently did.

Two takeaways for enterprises

The AI assistant is a second recipient. When you give an AI email assistant access to a mailbox, you have added a second security-relevant reader to every inbound message, and that reader has a different threat model than the human. The mail security tier was built around the human reader. If you treat the AI as "just a summarizer" and leave your controls at the gateway tier, the attacker-chosen summary is what the human sees.

The content plane and the instruction plane are not the same. This is the structural observation under the four hiding techniques. Any content channel the AI reads during a task — email body, résumé, invoice, support ticket, README, install guide, dependency metadata, MCP tool description, config file — is a potential instruction channel. The defender posture that holds up as that list grows is to watch what the AI does, not just to try to catch everything hidden in what it was given.

Barracuda is explicit that the campaign they're writing about is real and current, not a lab construct. The four hiding techniques are mechanical and defeasible. The five downstream behaviors — summary override, invoice BEC, résumé auto-10/10, chatbot configuration exfil, code-assistant credential line — are each measurable at the moment of action. The gap between "the AI read this" and "the AI did this" is where the boundary has to be rebuilt.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.