LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems

2026-05-22 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

LCGuard, a framework for safe KV-based latent communication in multi-agent LLM systems, addresses the challenge of sensitive information leakage through shared transformer key-value (KV) caches. These caches encode contextual inputs and intermediate reasoning states, creating an opaque channel for sensitive content. LCGuard learns representation-level transformations before cache artifacts are transmitted, preserving task-relevant semantics while reducing reconstructable sensitive information. The framework formalizes leakage operationally through reconstruction, leading to an adversarial training formulation. Empirical evaluations across Qwen3 (4B, 8B, 14B), Gemma-2-9B, and LLaMA (3B, 8B) models on AgentLeak, MAGPIE, and PrivacyLens benchmarks show LCGuard consistently reduces reconstruction-based leakage and attack success rates (ASR) by approximately 65-75% while maintaining competitive task performance compared to standard KV-sharing baselines.

Key takeaway

For Machine Learning Engineers deploying multi-agent LLM systems that share internal KV caches, you must address the inherent sensitive information leakage. LCGuard offers a principled solution by transforming KV representations, reducing attack success rates by 65-75% while preserving task performance. You should consider implementing LCGuard, especially its Full-System variant, to mitigate representation-level privacy risks and ensure safer, more efficient latent communication without relying solely on output-level safeguards.

Key insights

Latent communication via shared KV caches in multi-agent LLMs poses a privacy risk, which LCGuard mitigates through adversarial representation-level transformations.

Principles

• Latent KV communication creates high-capacity leakage channels.
• Output-level privacy methods fail to control latent leakage.
• System-level optimization is crucial for compositional leakage.

Method

LCGuard uses an adversarial learning framework to jointly optimize communication functions and reconstruction models. Communication functions transform KV representations to maximize reconstruction loss while preserving task utility.

In practice

• Implement LCGuard for KV-based multi-agent LLM systems.
• Optimize transformations system-wide for compositional leakage.
• Adjust "beta" to balance privacy and task utility.

Topics

• Multi-agent Systems
• Latent Communication
• KV Cache Security
• Adversarial Training
• LLM Privacy
• Representation Learning

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems — LCGuard secures latent KV cache communication in multi-agent LLMs by transforming representations to prevent sensitive data reconstruction.
• When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems — Embedding-based MAS defenses fail when attackers craft near-benign messages, necessitating token-level confidence signals for robustness.
• SAIGuard: Communication-State Simulation for Proactive Defense of LLM Multi-Agent Systems — Proactive simulation of multi-agent communication can detect and mitigate security risks before they propagate, preserving system utility.
• The Art of Mixology: Mixup-based Obfuscation for Privacy-Preserving Split Learning in Large Language Models — MIXGUARD enhances privacy and utility in LLM split learning through token- and representation-level mixup obfuscation.
• A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle — LLM agent memory's inherent malleability necessitates lifecycle-wide security governance for mnemonic sovereignty.
• Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems — Latent communication in LLM multi-agent systems reduces cost and information loss by exchanging continuous representations directly.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.