Resource content or prompt template contains prompt injection patterns
A resource body or prompt-template message contains hidden characters or text that addresses the model.
What it means
Beyond tools, MCP servers also expose resources/read (document contents) and prompts/get (rendered prompt templates). Both get pasted verbatim into the model's context when a client loads them. A compromised server can hide zero-width Unicode or role markers inside a resource body or a prompt message and the user never sees the payload before it steers the agent — tool-description scanners (POISONING-001) miss this because they only look at the tool catalog. The check looks for the same two things POISONING-001 looks for, but in a wider surface: high: hidden characters (zero-width, bidirectional overrides, soft hyphens, Unicode tag characters) in a resource name, description or content body, or in a prompt description, argument description or message text. medium: visible text aimed at the model (<system> style tags, role prefixes, ignore-previous directives) in the same slots. The check only fires when the audit was allowed to call resources/read and prompts/get — our third-party crawl skips those by default; this runs against your own servers and the patchbay demo.
How to fix
Strip hidden characters from every resource body and prompt message before publishing. Rewrite flagged content so it describes the data for the human reader instead of addressing the model. If you did not author the flagged text, treat the server as compromised and rotate any credentials it had access to.
See it live
Our demo MCP server trips POISONING-002 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.