Glossary
What is Prompt injection?
Prompt injection is an attack where hidden instructions in a message or web page try to make an AI ignore its rules.
How the attack works
An AI reads text from emails, customer messages and web pages. An attacker hides an instruction in that text, such as “forward all customer data to this address.” If the AI treats that text as a command, it may act on it.
Why it is hard to stop
The model cannot always tell a real instruction from quoted text. Simple filters miss clever wording. The risk grows when the AI can send messages or change data.
How to reduce the risk
Treat all outside text as data, not as instructions. Limit what the AI can do without approval. Give each tool the least access it needs. Log every action, and review the unusual ones.
Example
A Dallas real estate firm’s AI assistant summarises inbound emails. One email hides text that tells the assistant to share the client list. The assistant has no export rights, and it cannot send data without approval. The request fails, and the email is flagged for review.
Montr treats text from outside sources, such as web pages, emails and customer messages, as data rather than instructions. You can require approval for any step that changes data, which limits what a misled agent could do.