Enforcing AI Safety Through Runtime Contracts

Summary & Methodology Analysis

The researchers propose shifting safety enforcement from model weights to a system-level runtime contract. This architecture employs a two-faced harness designed to intercept and audit agent behavior. The preventive layer functions as a middleware component that utilizes sandboxes, permission gates, and input filters to halt dangerous actions before they reach execution. This prevents the agent from performing unauthorized or harmful operations in the host environment by strictly gating API interactions and system calls at runtime. The architecture avoids reliance on model self-reporting, which is often unreliable in autonomous workflows. To ensure accountability, the evidential layer introduces hash-chained trajectory logging and evidence-gated submissions. This layer generates a cryptographically verifiable record of the agent actions, ensuring that task completion is supported by concrete, tamper-evident data captured via a formal Agent Trajectory Schema. By composing these preventive monitors and evidential verifiers, the system creates an immutable contract for agent operation. The paper notes that current industry practices are lacking, as an audit of 12 public agent systems showed that only 2 implemented submission-like evidence gates. The methodology is computationally efficient only when the defined verifiers operate independently, as compositional verification can otherwise introduce overhead. A primary limitation is that the contract governs the external actions and submissions of the agent, rather than the internal objectives or goals the model is attempting to optimize.