Natural-Language Permission Verification in Multi-Agent Delegation

Summary

As AI agents increasingly delegate tasks to other agents, permission scopes are often expressed informally in natural language rather than as strict machine-readable policies. This research investigates how a delegated action can be automatically checked against the natural-language permissions granted upstream, so that a multi-agent system can catch privilege escalation and out-of-scope actions before they execute rather than after the fact.

Objectives

• Define a natural-language permission model that describes what an agent is allowed to delegate to sub-agents.
• Design a verification layer that checks whether a delegated task or action stays within the scope of the original permission grant.
• Evaluate whether an LLM-based verifier can reliably detect permission violations across multi-hop delegation chains.
• Compare the natural-language approach against formal, rule-based policy models on accuracy, false-positive rate, and usability.

Methodology

1. 01Review existing access-control models (RBAC, ABAC, capability-based systems) and how current multi-agent frameworks handle delegation.
2. 02Design a specification format for expressing permissions in natural language, with defined scope and delegation boundaries.
3. 03Implement a verification layer that intercepts delegation requests between agents and checks them against the granted permissions.
4. 04Build a benchmark of delegation scenarios covering legitimate delegations and permission-escalation attempts.
5. 05Run empirical evaluations measuring verifier accuracy, latency overhead, and robustness against adversarial phrasing.

Findings

Findings are still being collected as the study progresses. This section will be updated with results as the evaluation phase completes.

Publications

No publications yet — this section will be updated as the research is submitted or published.

Code

Repository link coming soon.