Extend multi-CaMeL security beyond hierarchical trees

Extend multi-CaMeL and its system-level control-flow-integrity and CaMeL-security guarantees from hierarchical tree-structured agent-as-tool systems to directed acyclic graphs, cyclic interactions, asynchronous message-passing systems, shared-state systems, and peer-to-peer agent architectures.

Background

The paper proves security guarantees for multi-CaMeL only in hierarchical agent-as-tool systems whose communication topology is a tree. In that setting, each non-root agent has one parent and communication proceeds through parent-to-child calls and responses, making provenance tracking and enforcement of the trusted-instruction/data distinction well defined.

More general architectures may involve multiple parents, retained conversational state, peer-to-peer communication, asynchronous message queues, shared state, or cycles. These settings lack the simple root-to-leaf information-flow structure used in the paper and may require new mechanisms for preserving provenance and enforcing system-level control-flow integrity and CaMeL-security.

References

Extending these ideas to DAG, cyclic, asynchronous, and peer-to-peer agent system topologies, and to stronger information-flow properties, remains open for future work.

— Can CaMeLs Talk? Securing Multi-Agent Systems Against Indirect Prompt Injection Attacks  (2610.05640 - Peters-Gill et al., 5 Oct 2026) in Section 6, Conclusion