A page can be useful reference material and still contain instructions an AI should not follow.

That is the practical problem behind indirect prompt injection: content from a page, document or tool result can try to redirect the system’s behavior.

Keep external content separate from trusted instructions. Limit tool permissions and set approval gates for consequential actions. These layers reduce risk; they do not guarantee safety.

What is the most consequential action your assistant can actually take?

More on proposals and actions: https://dash-squared.com/learn/ai-workflows/

#AISecurity #PromptInjection #ITOperations