AI RESEARCH
The Capability Paradox: How Smarter Auditors Make Multi-Agent Systems Less Secure
arXiv CS.AI
•
ArXi:2605.17480v2 Announce Type: replace Multi-agent systems extend large language models (LLMs) by decomposing tasks among specialized agents, but their distributed decision process creates new attack surfaces. We identify semantic hijacking, an attack in which harmful requests are concealed within domain-specific narratives and propagated to a Manager through Worker reports, without any syntactic injection primitives.