Exploiting Delegated Trust Between Agents
Orchestrators delegate tasks to sub-agents by passing instructions and context. Sub-agents typically have no mechanism to verify that the entity sending them instructions is actually their legitimate orchestrator. They operate on an implicit trust model: anything arriving on their input channel is treated as a trusted instruction. This trust can be abused from two directions — by impersonating the orchestrator to issue unauthorized instructions, or by compromising the orchestrator itself and using its trusted position to instruct sub-agents to perform actions outside their intended scope.
Abuse the Orchestrator's Trusted Position Through Injection
Compromise the orchestrator via prompt injection and use its privileged position to issue instructions to sub-agents it normally would not send:
If the orchestrator passes this instruction to the code agent without validating whether the user is authorized to issue code execution commands, the delegated trust from orchestrator to sub-agent is exploited. The user gains code execution capabilities they should not have by routing through the orchestrator's trusted relationship with the executor.
Exploit Scope Boundaries Between Agent Roles
Sub-agents are often scoped to specific task types but accept instructions from the orchestrator without checking whether the requested task falls within their intended scope. Test whether a research agent will execute tasks meant for the code or executor agent:
Exploit Trust Transitivity
In deep agent hierarchies, trust propagates through multiple levels. An orchestrator trusts sub-agent A, sub-agent A trusts sub-agent B, and so on. Compromising any link in the chain gives influence over everything downstream:
Test Whether Sub-Agents Validate Task Authorization
Send a task to a sub-agent that falls clearly outside its documented purpose to check whether it enforces scope:
An incoming connection on the listener confirms the sub-agent executed the out-of-scope instruction because it came from the orchestrator's trusted position. The sub-agent performed a network call that has nothing to do with its summarization role, demonstrating that delegated trust is not bounded by role scope.
References
-
OWASP Top 10 for LLM Applicationsowasp.org/www-project-top-10-for-large-language-model-applications (opens in new tab)
LLM08 Excessive Agency covering delegated trust exploitation
-
Anthropic — Building Effective Agentswww.anthropic.com/research/building-effective-agents (opens in new tab)
Orchestrator and sub-agent design patterns and trust model reference
Was this helpful?
Your feedback helps improve this page.