Why Static Config Checks Are Not Enough to Secure AI Agents With Tool Access

As LLM-based agents gain the ability to execute tools — such as shell commands, HTTP fetchers, and cloud SDKs — the risk shifts from harmful text output to harmful real-world actions performed with user credentials. Security researchers highlight four key abuse patterns: remote code execution via shell tools, server-side request forgery through HTTP fetchers, credential exfiltration via environment variables, and unauthorized tool calls triggered by prompt injection. Static configuration scanners, such as the open-source correctover-scan tool, can catch setup-level issues like plain-HTTP transports, hardcoded credentials, and overly broad permissions across common MCP config files. However, experts warn that static analysis cannot evaluate the specific arguments an agent constructs and passes to tools at runtime, which is where many real threats materialize. Runtime verification is therefore presented as a necessary complement to configuration hygiene for teams deploying agents with tool access.