Docker advocates for unified security layer beneath AI agent harnesses

In 1988, Norm Hardy identified the 'confused deputy' problem—software misusing its permissions—which remains relevant today as AI agents increasingly operate with similar authority. Docker argues that current per-harness security measures fail because agents can circumvent safeguards (e.g., using APIs when git pushes are blocked) and lack centralized governance across multiple models and harnesses. The company proposes a runtime layer beneath all harnesses to enforce uniform policies on code execution, tool usage, and credentials. Docker cites three industry trends driving this approach: the high cost of running frontier models for simple tasks, frequent shifts in top-performing AI models, and growing reliance on custom models trained with proprietary data. Recent research demonstrated risks when a GitHub issue manipulated a coding agent into leaking private repository contents via legitimate access channels—highlighting how existing guardrails fail when agents interact dynamically with their environment. A runtime security layer would maintain consistent boundaries regardless of model behavior changes or harness updates, centralize audit trails, and enable granular control without sacrificing automation. Docker CTO Tushar Jain will discuss this framework at the We Are Developers conference in San Jose.