{"@context":"https://schema.org","@type":"CreativeWork","@id":"https://froggit.ai/public/capsules/34bd5e2d-f4bc-4607-9f54-89736f3dde15","identifier":"34bd5e2d-f4bc-4607-9f54-89736f3dde15","url":"https://froggit.ai/public/capsules/34bd5e2d-f4bc-4607-9f54-89736f3dde15","name":"LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems","text":"# LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems\n\nSource: arXiv:2605.22786, submitted May 21, 2026.\nAuthors: Sadia Asif et al.\n\nThis paper studies safety risks in multi-agent LLM systems that share transformer key-value caches or other latent communication artifacts. It argues that KV caches can carry contextual inputs, intermediate reasoning state, and agent-specific information even when no sensitive text is explicitly sent.\n\nPublic takeaways:\n- Latent communication can improve coordination and preserve task-relevant information, but it creates an opaque channel for sensitive information leakage.\n- LCGuard treats shared KV caches as latent working memory and learns transformations before cache artifacts are transmitted between agents.\n- The paper defines leakage through reconstruction: a cache artifact is unsafe if an adversarial decoder can recover sensitive agent-specific inputs from it.\n- The reported evaluations reduce reconstruction-based leakage and attack success while preserving competitive task performance against standard KV-sharing baselines.\n\nWhy this matters for Forge: this is a useful public reference for agent safety design, especially when evaluating shared memory, multi-agent coordination, privacy boundaries, and hidden-state transfer between agents.\n\nLimitations: this is an arXiv preprint/paper record. Forge has verified the source identity and made the capsule answer-ready as a source-backed reference, but has not independently reproduced the experiments.\n\nSources:\n- https://arxiv.org/abs/2605.22786\n","keywords":["agent-safety","arxiv","cs.AI","cs.ET","cs.LG","cs.MA","free-public-reference","kv-cache","latent-communication","llm-security","multi-agent-systems","privacy","source-backed","public-reference"],"about":[],"citation":["https://arxiv.org/abs/2605.22786"],"isPartOf":{"@type":"Dataset","name":"Forge Cascade Knowledge Graph","url":"https://froggit.ai"},"publisher":{"@type":"Organization","name":"Forge Cascade","url":"https://froggit.ai"},"dateCreated":"2026-05-22T06:00:06.024000Z","dateModified":"2026-06-19T09:56:41.332000Z","isBasedOn":"https://arxiv.org/abs/2605.22786","additionalProperty":[{"@type":"PropertyValue","name":"trust_level","value":100},{"@type":"PropertyValue","name":"verification_status","value":"sources_verified"},{"@type":"PropertyValue","name":"provenance_status","value":"valid"},{"@type":"PropertyValue","name":"evidence_level","value":"primary_source"}]}