ActPlane: Programmable OS-Level Policy Enforcement for Agent Harnesses (arxiv.org)

A new initiative titled ActPlane has been introduced to enhance the safety and efficacy of AI agents operating in production environments. ActPlane addresses the challenges of enforcing operating system (OS) policies that govern AI behaviors, which often arise from ambiguously defined natural language. Traditional enforcement mechanisms, such as tool-call guardrails and OS sandboxes, either miss critical system actions or generate opaque errors that hinder AI decision-making. ActPlane overcomes these limitations by integrating policy enforcement directly into the OS kernel, enabling agents to declare policies that are upheld with semantic feedback. Significantly, ActPlane leverages an information-flow control (IFC) domain-specific language (DSL) to manage cross-event policies, which broadens the scope of enforceable actions beyond what conventional methods allow. Implemented using extended Berkeley Packet Filter (eBPF), ActPlane demonstrates improved policy compliance, particularly for indirect execution paths that previous interceptions failed to cover, achieving only a 1.9%-8.4% performance overhead. This innovation is noteworthy for the AI/ML community as it offers a robust framework for ensuring that AI agents adhere to safety standards while preserving operational efficiency.