AI Agents Should Propose Infrastructure, Not Mutate It (devgeist.com)

Recent discussions in the AI/ML community highlight the risks associated with current cloud vendor MCP (Managed Cloud Platform) servers that allow AI agents to alter infrastructure directly. This trend raises significant concerns about governance, as demonstrated by incidents where AI agents have accidentally deleted production databases or executed harmful commands through compromised integrations. The crux of the issue lies in the permissions granted to these agents, which can perform destructive actions without proper oversight, ultimately reversing advancements made through infrastructure-as-code (IaC) approaches. Experts advocate for a shift in how AI agents interact with cloud infrastructure, suggesting that these agents should be limited to proposing changes rather than executing them. This concept promotes a separation of powers, ensuring that an agent's authority to suggest infrastructure modifications is distinct from the authority to approve and execute those changes. By implementing a structured governance framework that transforms agent proposals into declarative artifacts subject to organizational policies, the industry could maintain oversight and reduce operational risks. This method not only preserves the core benefits of IaC—such as versioning and checkability—but also mitigates potential destructive outcomes emerging from AI-enhanced operations, safeguarding against the pitfalls of unchecked automation.