MirLabs / The lab / Prompt versus code
Planned
Prompt versus code
To prevent an agent's faults, how much does a security instruction in the prompt help, and how much does a guard in the code?
Result
No result yet.
Protocol
The same agent goes through the same traps in three setups: a plain prompt, a prompt loaded with security instructions, and the plain prompt with a guard in the code that refuses forbidden acts. We count the faults in each setup.
Limits
A guard only covers the acts it knows: the result will say what it blocks in these traps, not in all of them.