Distributing Security Controls Through Harness Engineering

Research Context and Approach

As AI coding agents become ubiquitous, security concerns have emerged as the primary barrier to their organizational adoption. While individual security controls exist, they are often difficult to deploy consistently across distributed engineering teams. Furthermore, many vendor-native security solutions lock organizations into specific ecosystems. This research investigates whether security controls can be effectively "harnessed"—embedded directly into the agent's execution environment—to provide a scalable, distributable security layer.

The author developed SHarD (Secure Harness Distribution), a custom harness built upon the Pi agent framework. The study utilized a 23-test suite derived from the OWASP Top 10 for Agentic Applications to evaluate four configurations: commercial agents with and without controls, a baseline harness, and the security-hardened SHarD harness.

Key Findings

The research successfully validated that three specific categories of security controls—OS sandboxing, skill scanning, and tool restriction—can be embedded into a harness and distributed via a single installation command. SHarD achieved a 100% adjusted score in the test suite, matching the efficacy of the best-configured commercial agents without introducing regressions.

Notably, the study found that model non-determinism leads to inconsistent security outcomes when relying solely on the agent's internal logic. By contrast, OS sandboxing provided a robust, deterministic boundary that successfully mitigated autonomous agent behaviors that would otherwise cross system boundaries. The author also identified that content protection mechanisms (such as AI gateways or guardrails) currently impose prohibitive overhead, making them unsuitable for inclusion in a lightweight, distributable harness at this time.

This work provides a practical blueprint for engineering teams to treat security as a first-class component of the agent harness rather than an afterthought. By moving security controls into the harness, organizations can ensure consistent protection across diverse developer environments regardless of the underlying model or vendor. This approach supports a "defense-in-depth" strategy that is both scalable and manageable, offering a path forward for organizations struggling to balance the rapid adoption of agentic AI with necessary risk mitigation.