When agents delegate to other agents, one compromised or malicious participant can pass injected instructions down the chain or talk a peer into lifting its own restrictions. Treat every message between agents as untrusted input and require authenticated agent identity.
Read in this order
- OfficialOWASP: Multi-Agentic System Threat Modeling Guide v1.0Start here: a method for threat modeling systems of cooperating agents.
- ArticleBrief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident (METR)How agents coordinated through shared state during a real incident, August 2026.
- VideoMulti-Agent Security Is Not Your Traditional Security | Cisco Live EMEA 2026Explains how shared memory and agent handoffs propagate compromise. 38 min.