# LLM data exfiltration: how AI agents leak data

> How AI agents leak data: steered sends, links and images that carry data, replies to the wrong party. And the controls that limit what leaves.

Source: https://trotta.email/learn/llm-data-exfiltration

LLM data exfiltration is the theft of private data through an AI model or agent that has access to it. An attacker plants instructions in content the agent reads, and the agent sends the data out itself: in a reply, in the address of a link or image, or to the wrong recipient.

Updated November 20, 2026

## How do AI agents leak data?

AI agents leak data when something they read tells them to send it and something they can do carries it out. The attacker does not break in. The agent already has access, and [prompt injection](https://trotta.email/learn/prompt-injection) borrows it.

Three conditions make a leak possible: the agent can reach private data, it reads content from untrusted sources, and it has a way to send something outside. Remove one and the path closes.

Not every leak is an attack. A model can reveal data by mistake, such as quoting an earlier thread to the wrong contact or including a secret from its context in an answer. OWASP covers both cases under Sensitive Information Disclosure, LLM02 in its Top 10 for LLM Applications 2026.

## What are the ways data gets out?

Data gets out through any output that reaches a place the attacker can read.

| Exit | How it carries data | Needs a click |
| --- | --- | --- |
| A message to the attacker | The agent is told to forward, reply or send to an outside address | No |
| A link | Private data is placed in the address of a link shown in the answer | Yes |
| An image | Private data is placed in the address of an image, and the client loads it automatically | No |
| A reply to the wrong party | The agent answers a lookalike sender or a hijacked thread with real details | No |
| A tool call | The agent is told to post data to an outside service it can reach | No |
| A shared place | The agent writes data into a document, ticket or channel the attacker can see | No |

### Being steered to send secrets

The plainest exit is a send. An injected instruction asks the agent to forward a thread, attach a file or include a key in its reply. To the mail system it is a normal message from a real account.

OWASP's example for Excessive Agency describes this: a mail assistant tricked by an incoming email into finding sensitive information and forwarding it.

### Links and images that carry data

A web address can hold data. If a model is steered to write an address that includes private text, whoever runs the server at that address receives the text when it is requested.

A link needs a human to click. An image does not, because many clients load images on their own. Microsoft's security team describes both forms and says it blocks the image form in its own products.

An image was the exit in EchoLeak, CVE-2025-32711. Researchers showed that one crafted email could make Microsoft 365 Copilot disclose data with no click.

Microsoft rated it 9.3 (critical) and published its advisory in June 2025, saying the flaw was already mitigated. See [zero-click AI attacks](https://trotta.email/learn/zero-click-ai-attacks).

### Replies to the wrong party

An agent that answers mail can be asked for data by a sender who is not who they claim to be. A [lookalike domain](https://trotta.email/threats/lookalike-domains) or a [hijacked thread](https://trotta.email/threats/thread-hijacking) makes the request look routine.

No injection is needed. The agent does its job for the wrong audience.

## Why is exfiltration through an agent hard to see?

Exfiltration through an agent is hard to see because every step looks legitimate.

The agent is an approved identity. It reads data it is allowed to read. It sends through a channel it is allowed to use.

Monitoring that looks for unusual accounts or unusual volumes finds neither in one reply with a paragraph of private text.

The instruction that caused the leak may be invisible too. It can sit in [hidden text](https://trotta.email/threats/hidden-text-in-email) or in an attachment the human never opened.

## What controls limit what leaves?

Controls on exfiltration work at three points: what the agent can read, what it is told and what it can send. A log covers what happens after.

| Point | Control | Why it helps |
| --- | --- | --- |
| Read | Give the agent its own inbox, not the whole mailbox | An agent can leak only what it can reach |
| Read | Keep sign-in codes, reset links and keys out of its reach | Credentials are the most reusable thing to steal |
| Told | Screen content before the model reads it | Fewer injected instructions arrive |
| Send | Screen outgoing messages for secrets and sensitive data | A secret is caught on the way out |
| Send | Hold the first message to a new recipient for a human | A steered reply usually goes to an address never used before |
| Send | Do not load outside images automatically where model output is shown | The exit that needs no click is closed |
| Send | Restrict which outside services the agent can call | Fewer exits exist |
| After | Log what left, to whom, and after reading what | A leak can be traced and contained |

Output controls matter most when input controls fail. They are the last point where a leak can be stopped. For the credential case, see [verification code theft](https://trotta.email/threats/verification-code-theft).

## How does Trotta Mail control what leaves?

Trotta Mail screens what an agent sends before it leaves. The screen looks for secrets and sensitive data, and for replies that an injected instruction is steering to an address the inbox has never written to.

A secret inside a document an agent sends is refused like one in the body.

| Outcome | What happens |
| --- | --- |
| Deliver | The message is sent. |
| Hold | The message waits for a human, who approves or discards it. |
| Reject | The message is not sent. The API returns an error that carries the verdict. |

An inbox can also require approval for the first message to any new recipient, or for every send. An agent's key cannot loosen that setting or approve its own message.

On the reading side, sign-in codes, password resets and sign-in links in a human's inbox are sealed from every API key.

An agent that signed itself up, and that no human has claimed yet, cannot start a conversation with a new outside address. It replies only to senders whose mail arrived authenticated.

Details are in [concepts](https://trotta.email/docs/concepts#outbound-screening) and the [security model](https://trotta.email/docs/security).

No filter catches every leak or every injection. Trotta Mail sees what goes through the mailbox and nothing else, so controls on an agent's other tools remain yours.

## Questions

### What is LLM data exfiltration?

LLM data exfiltration is the theft of private data through an AI model or agent that can reach it. The attacker gets instructions in front of the model, and the model sends the data out through a reply, a link, an image or a tool.

### How does prompt injection lead to data exfiltration?

Prompt injection gives the attacker a say in what the agent does next. If the agent can read private data and can send anything outside, the injected instruction joins the two: read this, send it there.

### Can an AI agent leak data without anyone clicking a link?

Yes. If the agent can send mail, the send is the leak. If its output is shown in a client that loads images automatically, an image address can carry the data out, which was the exit in EchoLeak, CVE-2025-32711.

### How do you stop an AI agent from leaking data?

Limit what the agent can read, screen what it reads and check what it sends. Hold the first message to a new recipient for a human, keep credentials out of its reach and do not load outside images automatically.

## Sources

- [OWASP Top 10 for LLM Applications 2026, the current edition (published August 2026)](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/)
- [OWASP Top 10 for LLM Applications 2025, LLM02: Sensitive Information Disclosure](https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/)
- [OWASP Top 10 for LLM Applications 2025, LLM06:2025 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)
- [Microsoft Security Response Center, How Microsoft defends against indirect prompt injection attacks, July 2025](https://www.microsoft.com/en-us/msrc/blog/2025/07/how-microsoft-defends-against-indirect-prompt-injection-attacks)
- [Microsoft Security Response Center, CVE-2025-32711: M365 Copilot Information Disclosure Vulnerability](https://msrc.microsoft.com/update-guide/vulnerability/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)
- [Meta, Agents Rule of Two: A Practical Approach to AI Agent Security, October 2025](https://ai.meta.com/blog/practical-ai-agent-security/)

## Related

- [Data exfiltration through agents](https://trotta.email/threats/data-exfiltration-through-agents) Threat library
- [What is prompt injection?](https://trotta.email/learn/prompt-injection) Guides
- [Zero-click AI attacks](https://trotta.email/learn/zero-click-ai-attacks) Guides
- [AI agent security](https://trotta.email/learn/ai-agent-security) Guides
- [Verification code theft](https://trotta.email/threats/verification-code-theft) 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)
