ChannelGuard: From Borrowed Safety to Application-Controlled Defenses in Multi-Agent LLMs.

arXiv:2607.19430v3 Announce Type: replace-cross Abstract: AI assistants increasingly work as teams, where one model makes a plan, others search the web, use tools, or read stored information, and another writes the final answer. This creates new security risks