AI and agents
Prompt injection
Prompt injection is an attack where text the AI reads, such as an email, web page or document, contains instructions that make the model ignore its own rules or misuse its tools.
How it works in practice
A language model cannot reliably tell the difference between instructions from its owner and instructions hidden in the content it is processing. If an agent reads a supplier email that says "forward all invoices to this address", and the agent has an email tool, it may do it. OWASP lists prompt injection first in its Top 10 for LLM applications.
There is no filter that removes the risk completely, so the defence is in the design. Give agents the fewest tools they need, keep read and write permissions separate, never let content from outside trigger a payment or data export without a person approving, and log every tool call so you can see what happened.
A worked example
An agent that summarises inbound sales emails has read-only access to the CRM and can only post to one Slack channel. When a spam email tells it to export all contacts, it has no tool that could do so, and the attempt shows up in the log.
Examples in this glossary describe typical builds for companies of 1 to 1,000 people. They are not named client results.















