Weighing the Envelope: Component Weights and Arbitration in Single-LLM-Agent Envelope-Layer Defenses
Weighing the Envelope: Component Weight and Arbitration in Envelope-Layer Defenses for a Single LLM Agent
The paper measured the effect of labels, memory, tool roles, and the backend on invocation decisions in the envelope-layer defense of a single LLM agent, finding that tool-side mandatory label verification greatly reduces semantic overreach, but strong success memory pulls it back, and memory-side downgrade statements can return results to the level of labels alone. Prior work targeted multi-agent A2A envelopes; this paper fills in the case where MCP tools and long-term memory coexist. The threat model assumes an attacker pre-plants malicious tools and poisons long-term memory, inadvertently triggered by benign user requests. Mechanistically, labels work through mandatory tool checks plus default deny; memory authorization wording weakens the defense with dose, and position and multi-turn pressure do not change the direction of the defense. Tool roles divide the label effect into substitution, suppression, and boundary failure; low penetration does not equal safety, and local leakage is the main residual risk. The authors point out that envelope labels are not a switch but a measurable, reversible weight constrained by the backend and tool roles, and can only serve as soft gating and a risk signal within broader defenses, not as a replacement for permissions, sandboxes, or human review.