Study
Agentic AI · Agent safety
The main risks when agents act (prompt injection through tool outputs, excessive permissions, unconfirmed side effects) and simple defences: separating data from instructions, allow-lists, argument validation and human confirmation.
- Anything an agent reads (web pages, emails, files) can contain instructions aimed at it; that's indirect prompt injection.
- Least privilege plus human confirmation for side effects limits the damage when a model is fooled.