Agent Loyalty

Whose side is the agent on when receiving instructions from multiple parties? By default, the model helps “whoever speaks”, but across the negotiation table sits an opponent with opposing interests. Spectrum of failures: excessive honesty (reveal the minimum price to the opponent) ↔ excessive suspicion (refuse even the owner). Prompt injection is essentially an attempt at defection. A set of loyalty rules in the prompt: protect the owner’s private information including the fact of its existence; do not list reasons for refusal; private position ≠ public; only execute explicit instructions from the owner; withstand pressure.

Related: [Deadly Triad], [Prompt Injection]