The Agentic Web Has a Trust Problem — And It's Already in Production

What Happened Researchers audited 2,214 real-world Model Context Protocol (MCP) servers - the emerging standard for connecting LLMs to external tools - and found that 9.93% of tool descriptions don't match their underlying code. Since LLM agents trust these natural-language descriptions to decide what to execute, the mismatch ("Description-Code Inconsistency") opens a path from benign bugs to stealthy malicious behavior. The team built DCIChecker, an automated scanner, hinting at a nascent category of agent-tool verification tooling.