Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Data exfiltration (AI agents)

Data exfiltration, for AI agents, is when an attacker tricks an agent into reading private data it can access and sending it out, through a web request, an email, an image link or another channel the attacker controls.

Also called: exfiltration, lethal trifecta

By Divya Sharma · Checked Nov 15, 2026 · How we review

Exfiltration is an old security word for moving data out of a system without permission. Agents make it easier, because they can read private data and also talk to the outside world, and they follow instructions found in the text they process.

The lethal trifecta

Simon Willison describes the risk as three features in one agent: access to private data, exposure to untrusted content, and a way to communicate externally. If all three are present, a prompt injection hidden in a web page, document or issue can tell the agent to gather secrets and send them to the attacker. Remove any one of the three and this attack path closes.

How data leaves

The channel is often something that looks harmless: a URL with the data packed into its query string, a Markdown image the chat client loads automatically, a commit pushed to a public repository, an outgoing email, or an argument passed to a tool. Willison notes that vendors usually fixed reported cases by closing the specific channel used to send the data out.

Example

A coding agent with access to a private repo is asked to fix a bug from a public issue. The issue text contains hidden instructions to read the `.env` file and include its contents in a link. If the agent can make network requests, the keys leave with that request.

Defences

Limit what the agent can reach with narrow agent permissions. Run it in an agent sandbox with outbound network access blocked or restricted to an allowlist. Do not render remote images from model output, and require human approval for actions that send data out. Tool poisoning is one common way the injected instructions arrive.

Related terms

Prompt injection
Prompt injection is an attack where text fed to a language model, typed by a user or hidden in a web page, file or tool result, changes the model's behaviour in ways its developer did not intend.
Agent sandbox
An agent sandbox is an isolated environment, such as a container, virtual machine or operating-system restriction, in which an AI agent runs code and commands so it cannot reach files, networks or data outside set limits.
Agent permissions
Agent permissions are the rules that decide which tools, files, commands, networks and accounts an AI agent may use, and which actions need a human's approval first, so a mistaken or manipulated agent can do limited damage.
Tool poisoning
Tool poisoning is an attack where a malicious tool, often an MCP server, hides instructions in its tool description or metadata; the AI model reads them as guidance and may leak data or misuse other tools without the user noticing.
Human-in-the-loop
Human-in-the-loop (HITL) is a design in which an AI system pauses at chosen points for a person to approve, correct or reject its next step, so risky or uncertain actions are not taken automatically.

Sources