Multi-step and multi-agent attacks
One compromised agent passes the attack to others: a low-privilege agent persuades a privileged one to act, and failures cascade through a system of agents.
- ASI07:2026 Insecure Inter-Agent Communication
- ASI08:2026 Cascading Failures
- ASI10:2026 Rogue Agents
- AML.T0143 LLM Prompt Propagation
How it works
- 1In multi-agent systems, agents send each other tasks and results as text.
- 2An agent that trusts another agent's output inherits whatever that output carries, including instructions planted upstream.
- 3A privileged agent acting on a request from a compromised one is a confused deputy: the action is authorised, the intent is not.
Real cases on the map
- CriticalOpenClaw: Plugin subagent routes could bypass gateway authorization with synthetic admin scopesopenclaw
- HighOpenClaw: Leaf subagents could steer sibling sessions across sandbox boundariesopenclaw
- High`OpenClaw: session_status` let sandboxed subagents access parent or sibling session stateopenclaw
- MediumOpenClaw's ACP child sessions inherit subagent security envelope constraintsopenclaw
- MediumOpenClaw: Gateway Plugin Subagent Fallback `deleteSession` Uses Synthetic `operator.admin`openclaw
- MediumOpenClaw leaf subagents can bypass controlScope restrictions to send messages to child sessionsopenclaw
Published advisories whose own titles describe this technique.
What to watch for
- Requests between agents that exceed the sender's own permissions
- A failure or anomaly in one agent followed by unusual actions in another
- Agents acting without a traceable user request
How to defend
- Authenticate agents to each other and check each request against the sender's rights.
- Do not pass one agent's output to another as instructions.
- Contain failures: limit what one agent's output can trigger downstream.