# Zero-click AI attacks and EchoLeak (CVE-2025-32711)

> A zero-click AI attack needs no click because the assistant reads the message itself. EchoLeak (CVE-2025-32711) explained, and what changes for defenders.

Source: https://trotta.email/learn/zero-click-ai-attacks

A zero-click AI attack is an attack that succeeds without the target clicking, opening or approving anything, because an AI assistant reads the malicious content itself. The assistant, acting with a human's access, follows instructions hidden in a message or file and discloses data or takes an action.

Updated November 20, 2026

## What makes an AI attack zero-click?

An AI attack is zero-click when the assistant does the step a human used to do. Classic phishing needs a human to click a link, open a file or type a password.

A zero-click AI attack needs none of that. The assistant reads the message as part of its normal work, and the reading is the attack.

The technique underneath is [indirect prompt injection](https://trotta.email/learn/indirect-prompt-injection): instructions placed in content a model will process. Zero-click describes the delivery. No human decision sits between the attacker's message and the model.

|  | Classic phishing | Zero-click AI attack |
| --- | --- | --- |
| Target | A human | The assistant that works for the human |
| What must happen | A click, a download or a reply | The assistant reads the content |
| What the human sees | A suspicious message, if they look | Often nothing unusual |
| What does the damage | A stolen password, malware or a payment | The assistant's own access to data and tools |
| What training can do | Teach the human to pause | Little: no human is asked |

## What was EchoLeak (CVE-2025-32711)?

EchoLeak is the name given to CVE-2025-32711, a flaw in Microsoft 365 Copilot. Researchers showed that one crafted email could make the assistant disclose data with no click. Microsoft rated it 9.3 (critical).

The CVE record describes it as AI command injection that lets an unauthorized attacker disclose information over a network. Its severity rating records that no interaction from the victim is needed.

Microsoft published its advisory in June 2025. The advisory says the flaw was already fully mitigated by Microsoft and that customers needed to take no action. It lists the flaw as not exploited.

A published case study describes the chain in broad stages:

1. An attacker sends an email to someone in the organization. It is worded so that it does not read as an instruction to a machine.
2. Later, the human asks the assistant an ordinary question. The assistant retrieves related content, and the email is among it.
3. Instructions in the email cause the assistant to put internal data from its context into the address of an image in its answer.
4. The client loads the image automatically. The request carries the data to a server the attacker controls. Nobody clicked.

Each stage got past a separate safeguard, including an injection filter, link redaction and a content security policy.

The assistant was never the attacker's account. It acted with the victim's access, on the victim's data. Untrusted content from outside caused the model to reach trusted data inside.

## Why does EchoLeak matter beyond one product?

EchoLeak matters because the pattern is general. Any assistant that reads content from strangers, holds private data and can produce output that reaches the outside is exposed to the same kind of chain.

The specific flaw was fixed. The class of flaw remains.

The widest path is email. A mailbox accepts messages from anyone, and an assistant connected to it reads them without being asked. OWASP ranks prompt injection as LLM01, the top risk, in its Top 10 for LLM Applications 2026.

## What changes for defenders?

Zero-click attacks move the point of defense from a human's judgment to the moment before the model reads. Four things change.

- **Awareness training does not reach this.** No human is asked to decide anything, so there is no decision to train.
- **After delivery is too late.** A message flagged once it sits in the mailbox may already have been read by the assistant. Screening has to finish before the content is readable.
- **The reader is a model.** Filters tuned to what fools a human miss text written for a machine, including text a human cannot see.
- **Output is an exit.** Links, images and replies the assistant produces need their own checks, because that is how data leaves. See [LLM data exfiltration](https://trotta.email/learn/llm-data-exfiltration).

The controls that follow are layered. Each one has a limit, which is why none stands alone.

| Layer | What it does | What it does not do |
| --- | --- | --- |
| Screen content before the model reads it | Stops or marks many injections at the entry | Catch every rewording |
| Give the model what a human could see | Removes text hidden by styling or comments | Stop an instruction in plain sight |
| Limit what the assistant can reach | Bounds what one message can expose | Prevent the injection |
| Block automatic loading of outside images | Closes the exit that needs no click | Stop a human from clicking a link later |
| Screen what the assistant sends | Catches secrets and replies steered to strangers | Help if data leaves another way |
| Human approval for risky actions | Puts a decision back where one was missing | Scale to every message |

## Where does Trotta Mail fit?

Trotta Mail applies the first layer at the mailbox. Each message is screened before it is delivered, while the sending server is still connected.

There is no moment when unscreened mail is readable, by an agent or by a human. See [pre-delivery screening](https://trotta.email/learn/pre-delivery-email-screening).

Dangerous mail is held outside the inbox until a human releases it. An agent's key sees that a held message exists and why, and nothing its sender wrote.

Delivered mail reaches an agent as the safe view: hidden text removed, instructions aimed at a model replaced by a note, and links that are not clean defanged. Text inside attached documents is screened too.

What an agent sends is screened before it leaves, for secrets and for replies steered to an address the inbox has never written to. The rules are in the [security model](https://trotta.email/docs/security).

No filter catches every injection. A reworded message with no technical tell can still be delivered. Limits on what the agent can reach and send are what bound the damage when that happens.

## Questions

### What is a zero-click AI attack?

A zero-click AI attack needs no action from the human target. An AI assistant reads malicious content on its own, follows instructions inside it and leaks data or acts, using the access it was given.

### What was EchoLeak?

EchoLeak is CVE-2025-32711, a flaw in Microsoft 365 Copilot that Microsoft rated 9.3 (critical). Researchers showed that one crafted email could make the assistant disclose internal data with no click. Microsoft's advisory, published in June 2025, says the flaw was already mitigated.

### Was EchoLeak used in real attacks?

Microsoft's advisory lists the flaw as not exploited and says customers did not need to take any action. Security researchers reported it, and Microsoft fixed it in the service.

### Does security awareness training stop zero-click AI attacks?

No. Training helps a human pause before a click, and a zero-click attack asks no human for anything. The defense has to sit before the model reads the content, and around what the assistant can reach and send.

### Is the zero-click risk limited to Microsoft 365 Copilot?

No. The specific flaw was in one product and is fixed. The pattern applies to any assistant that reads content from strangers, holds private data and can send something out.

## Sources

- [Microsoft Security Response Center, CVE-2025-32711: M365 Copilot Information Disclosure Vulnerability](https://msrc.microsoft.com/update-guide/vulnerability/CVE-2025-32711)
- [CVE record, CVE-2025-32711 (published June 11, 2025)](https://www.cve.org/CVERecord?id=CVE-2025-32711)
- [NIST National Vulnerability Database, CVE-2025-32711](https://nvd.nist.gov/vuln/detail/CVE-2025-32711)
- [EchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System, arXiv:2509.10540](https://arxiv.org/abs/2509.10540)
- [OWASP Top 10 for LLM Applications 2026, the current edition (published August 2026)](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/)

## Related

- [Indirect prompt injection](https://trotta.email/learn/indirect-prompt-injection) Guides
- [LLM data exfiltration](https://trotta.email/learn/llm-data-exfiltration) Guides
- [What is pre-delivery email screening?](https://trotta.email/learn/pre-delivery-email-screening) Guides
- [Prompt injection by email](https://trotta.email/threats/prompt-injection-by-email) Threat library
- [Agent hijacking](https://trotta.email/threats/agent-hijacking) Threat library
- [Security model](https://trotta.email/docs/security) Docs

Give your agent an inbox of its own. Free for two inboxes. No card.

[Get an inbox](https://trotta.email/signup) [Read the docs](https://trotta.email/docs)
