Formal Policy Enforcement for Real-World Agentic Systems
Researchers from arXiv have introduced a new framework for enforcing security policies in agentic AI systems, addressing the limitations of current natural-language prompt-based methods. The proposed approach, realized in a system called FORGE, treats policy enforcement as a cross-cutting concern using aspect-oriented programming. Instead of relying on an agent's reasoning for compliance, the framework employs Datalog to define policies over abstract predicates describing execution context. An observability service maintains these predicates under a formal assume/guarantee contract, while a reference monitor enforces decisions at every relevant action. This method ensures deterministic enforcement and supports static analysis for verifying policy intent, eliminating ambiguities inherent in natural language. The framework is particularly effective for multi-agent systems where causal history matters. Evaluation through three case studies—prompt injection defense, pharmacovigilance approval workflows, and customer service organizational policies—demonstrates that FORGE can enforce complex policies without modifying underlying agents, offering a robust solution for secure real-world AI deployments.
Wire timeline
Formal Policy Enforcement for Real-World Agentic Systems
Researchers from arXiv have introduced a new framework for enforcing security policies in agentic AI systems, addressing the limitations of current natural-language prompt-based methods. The proposed approach, realized in a system called FORGE, treats policy enforcement as a cross-cutting concern using aspect-oriented programming. Instead of relying on an agent's reasoning for compliance, the framework employs Datalog to define policies over abstract predicates describing execution context. An observability service maintains these predicates under a formal assume/guarantee contract, while a reference monitor enforces decisions at every relevant action. This method ensures deterministic enforcement and supports static analysis for verifying policy intent, eliminating ambiguities inherent in natural language. The framework is particularly effective for multi-agent systems where causal history matters. Evaluation through three case studies—prompt injection defense, pharmacovigilance approval workflows, and customer service organizational policies—demonstrates that FORGE can enforce complex policies without modifying underlying agents, offering a robust solution for secure real-world AI deployments.
cs.AI updates on arXiv.org