Research · curated 1 Oct 2026
Share-Borne AI Virus: Memory-Hopping Attacks Across LLM Agents
First reported arxiv.org
Coverage timeline
Single-source research — first reported, latest, and curated coincide.
Why it matters
Artifact-mediated propagation shows that persistent memory plus shared files can turn independent LLM assistants into carriers of a self-replicating attack that outlives any single interaction, a worm-like failure mode defenders must anticipate as memory-equipped agents proliferate.
The research paper "Share-Borne AI Virus: Memory-Hopping Attacks Across LLM Agents" (arXiv:2609.35576) introduces artifact-mediated propagation, in which adversarial content planted in a shared artifact (e.g. a report) is stored in an LLM assistant's persistent memory, reproduced in subsequently created artifacts, and acquired by other assistants that later read them. Using simulated temporal human-agent environments, the authors measure how such attacks survive hand-offs and spread, reporting propagation chains reaching 60-80% of agents across up to eight hops, with accompanying code published on GitHub.