In plain English
Models follow text and cannot always tell your rules from text they read. That allows prompt injection. The worst case is private data, outside content and a way to send data out, all in one system. Remove one. Use least access, checks in code, human approval and regular attack tests.
- Prompt injection
- Hidden instructions in text the AI reads.
- Jailbreak
- A trick that makes a model ignore its safety rules.
- Least access
- Only the permissions the job needs.
Picture it
Defend in layers
Least access
Only what the job needs
Treat outside text as data
Label it, do not obey it
Checks in code
Allow lists and argument checks