Common attacks

Prompt injection

Hiding instructions in text, a web page or a file so that an AI assistant reads them and obeys them as if they came from you.

In plain words

It is a note slipped into a stack of papers your assistant reads: if the assistant cannot tell your instructions from the note, it follows the note.

When it affects you

Anyone who lets an AI assistant read emails, web pages, documents or tool output, and especially those who give it permission to act, such as sending messages or running commands.

What to do today

  1. Give an AI agent only the access it needs, and no keys or production data it can do without.
  2. Ask for confirmation before it takes actions that cannot be undone.
  3. Treat anything an agent reads from outside, such as a web page or an email, as untrusted.

← All termsNo. 21 / 32

Last reviewed: 2026-10-27. Written for non-experts. Check an advisory or your vendor for the details of a specific flaw.

Get the free runbooks

One-page security checklists for founders who build with AI. You get a confirmation email first, then the runbooks.

Send me the runbooks