Show HN: AgentWard – After an AI agent deleted files, I built a runtime enforcer (github.com)

AgentWard has been introduced as an open-source runtime enforcer designed to enhance permission control for AI agents that interact with various tools. Traditional methods, such as static scanners and prompt-based rules, have proven inadequate as they can be bypassed or simply identify risks without enforcing security measures. AgentWard fills this critical gap by functioning as a transparent proxy between AI agents and their tools (like MCP servers and APIs), allowing for real-time policy evaluation and enforcement of access controls. This means that, rather than relying on natural language instructions that can be easily ignored, AgentWard enforces compliance through code-level restrictions that remain out of reach of the AI's context. The significance of AgentWard to the AI/ML community cannot be overstated, especially as AI agents are increasingly given access to sensitive resources such as emails, calendars, and databases. By ensuring that every tool call is meticulously evaluated against a defined policy at runtime, AgentWard protects against potentially harmful actions, such as unauthorized file deletions or access to sensitive data. With features like smart-default policy generation, compliance audit trails, and various operational modes, AgentWard not only enhances security but also encourages responsible AI deployment in environments where agent capabilities pose significant risks.