Why it matters

Indirect PI is major agent threat. Defenses layered.

Advertisement

The architecture

Isolate untrusted content.

Signal boundaries clearly.

Verify output.

Indirect PI defensesIsolate contenttrust boundarySignal roleuser vs dataVerify outputbefore actionMultiple defenses needed; no single technique bulletproof
Indirect PI def.
Advertisement

How it works end to end

Provider role separation.

XML tags around untrusted.

Instruction hierarchy models.

Output verification before action.