Agent Safety as a Runtime Contract: A New Paradigm for AI Harness Design

A recent paper on arXiv (2608.11274) posits that ensuring AI safety for autonomous agents should occur during runtime through harnesses rather than through training. The authors argue that existing training techniques, such as RLHF, DPO, and Constitutional AI, are inherently inadequate for agents that can execute code, alter files, communicate, and adjust databases. They introduce a dual-runtime contract: a preventive aspect that restricts harmful actions via sandboxes, permission gates, output filters, and trajectory monitors, and an evidential aspect that mandates verifiable evidence of positive actions, linking task submissions to concrete proof like test runs, log captures, file diffs, and citation grounding. The paper supports its claims with four lines of public evidence, including a survey of 52 documented safety incidents involving AI agents and LLMs, as well as an audit of 31 uncontested false completions, with detailed protocols and data available in supplementary JSON files. This work signifies a notable shift in perspectives on AI agent safety, highlighting the runtime environment's critical role over training interventions.