Tool annotations contradict the tool name
A tool with a destructive-sounding name advertises destructiveHint:false, or a write-shaped name advertises readOnlyHint:true.
What it means
MCP 2026-07-28 added an `annotations` object on tools (`readOnlyHint`, `destructiveHint`, `idempotentHint`, `openWorldHint`, `title`). Clients use these hints to decide whether to show a confirmation prompt, whether to allow the tool into an auto-approve set, how to render the tool in a catalog. The spec and the official MCP blog are blunt: these are a UX layer, not a security layer — "clients MUST consider these untrusted unless the server is trusted". The check catches the specific case where a server is lying with the hints. A tool named `delete_everything` that sets `destructiveHint: false` has disabled the UX confirmation the hint was meant to govern. A tool named `create_user` that sets `readOnlyHint: true` has the same effect. Short list of unambiguous verbs (delete/drop/send/transfer/pay/charge, etc.) keeps false positives low. Ambiguous names are not flagged.
How to fix
Set annotations that match the tool's real behaviour. If a destructive tool was intentionally marked non-destructive to skip a confirmation dialog, either rename the tool (so the hint matches the name) or leave the hint absent (clients default to the safe side when annotations are missing). Client-side, do not take annotations on faith: validate them against the tool name or input schema before using them to skip user prompts.
See it live
Our demo MCP server trips ANNOTATIONS-001 on purpose. Scan it and you'll see this exact finding with evidence and fix text.
npx -y @patchbay-nebius/agent audit https://demo.patchbay-mcp.dev/mcpRun the check against your own server:
npx -y @patchbay-nebius/agent audit <url-or-directory>No install: scan an MCP server in the browser and get a shareable report.
Disagree with this finding on your server? Open an issue with the --json output. False-positive reports are the most useful feedback.