AI & ML
Weighing the Envelope: Component Weight and Arbitration in Envelope-Layer Defenses for a Single LLM Agent
arXiv cs.CR0 views
arXiv:2611.23844v1 Announce Type: new Abstract: The envelope layer is the channel through which malicious content enters an LLM agent. Prior work established label and memory effects on multi-agent A2A envelopes, but the weight structure when MCP tools and memory coexist in a single agent remains unclear. Our threat model assumes an attacker pre-plants a malicious tool and poisons long-term memory; a benign user request inadvertently triggers it, and the attacker never enters the conversation. Following GDA Measurement and A/B/C/D, we measure how labels, memory, tool roles, and the backend shape invocation decisions across label-memory conflict, tool roles, robustness, and tool generalization. A tool-side forced-verification label substantially reduces semantic breach, but strong success memory pulls it back; a memory-side downgrade statement returns the outcome to the level of the label alone. The label works through mandatory implementation inspection plus default denial, not length or format; memory authorization wording erodes it with dose, while position and multi-turn pressure do not change the defense direction. Tool roles divide label effects into substitution, suppression, and boundary failure: low penetration is not safety, local leakage is the main residual risk, and a second malicious tool shows the same direction. Envelope labels are not a switch but a measurable, reversible weight constrained by backend and tool role, acting as a soft gate and risk signal within a broader defense rather than a substitute for permissions, sandboxing, or human review.
Source:https://arxiv.org/abs/2611.23844PreviousPlug-and-Play Multi-Licensee Robust Watermarking for Mixture-of-Experts LLMsNextSmoothing and Learning for Movement-Robust RSS-Based Spoofing Detection
Related posts
Who called me about stolen ID and loans? Handle it like an unauthenticated security alert
0dev.to
Gemini 4 Models Compared on Reasoning, Coding, and Context
0dev.to
Beyond Frozen Weights: How Test-Time Training (TTT) Turns Inference into Live Learning
0dev.to

Why Turn-Based Voice AI Feels Broken: Building a 52ms Barge-In Loop (HardCall)
0dev.to
Harnesses! The other kind
0dev.to
Apple may be developing its own AI agent that runs your Mac
0TechCrunch