Explainer

The lethal trifecta: when one agent can read, be told and send

An agent that can read private data, takes in content from outside, and can send data out has everything an attacker needs. Take away any one of the three and the attack falls apart.

By the NoFly team

Three ingredients

Simon Willison named the pattern in June 2025. An agent becomes exploitable when a single run has all three of these at once:

  • Access to private data: a mailbox, a repository, a customer database, a file share.
  • Exposure to untrusted content: an email, a web page, a ticket, a pull request, a tool result written by someone else.
  • A way to send data out: an HTTP request, an email, a comment on a public issue, an image link the client will load.

The untrusted content carries an instruction. The agent follows it, reads the private data and sends it out through the third ingredient. No single step looks like an attack.

Why filters are not enough

Prompt-injection filters try to spot the instruction in the text. They help, but they are judging language, and a new phrasing gets through. The trifecta is about capabilities, so the durable fix is making sure no single run holds all three without someone noticing.

Where it has already happened

One email turned Microsoft 365 Copilot into a data leak. A public GitHub issue led an agent to copy private repository data into a public pull request. A support ticket led an agent holding a Supabase service key to paste secrets back into the ticket. Each one had all three ingredients.

Breaking the triangle

  • Give each agent only the data its task needs, and nothing it can read by default.
  • Treat content from outside as data, never as instructions, and screen it before the agent acts on it.
  • Limit where each agent may send data, and follow sensitive values from read to send.
  • When a run really needs all three, put a person in front of the step that sends.

How NoFly helps

  • Discover

    Each agent’s reach is mapped: the data it can read, the content it takes in and the places it can send to, so an agent holding all three stands out.

  • Protect

    Content from outside is screened before the agent acts on it, destinations are allow-listed per agent, and secrets and personal data are followed from where they were read to where they were sent.

  • Explain

    If an agent does act, the verdict names the passage that caused it, so the fix goes to the source rather than the symptom.

What to do this week

This week3 actions
  1. List the agents that can read private data and also read content from outside.
  2. For each one, write down every way it can send data out.
  3. Remove one of the three wherever the task allows, starting with outbound access.
Keep reading

More from the blog.

Find out what your AI is allowed to do today.

Book a call and we will map your agents, MCP servers and keys with you, or start on one laptop with the free scan.

Book a demo
npm i -g @nofly/agent