Skip to content
InfoResearchPreprintLLM-specific

Weighing the Envelope: Component Weight and Arbitration in Envelope-Layer Defenses for a Single LLM Agent

Published
Record updated
View JSON

Summary

This paper measures how labels, poisoned long-term memory, tool roles and the backend shape whether a single LLM agent invokes a malicious pre-planted tool triggered by a benign user request. A tool-side forced-verification label substantially reduces semantic breach, but strong success memory pulls the effect back, while a memory-side downgrade statement returns the outcome to the level of the label alone. The authors conclude that envelope labels act as a measurable, reversible weight, a soft gate and risk signal within a broader defense rather than a substitute for permissions, sandboxing, or human review.