patchbay audit
highPOISONING-003runtime-verified

Tool call result contains prompt-injection patterns

A tool whose annotations look safe returns content that addresses the model.

What it means

MCP tools can return arbitrary text via tools/call. Clients paste that text into the model's context, same as a resource body or a prompt message. POISONING-001 scans tool metadata, POISONING-002 scans resources/read and prompts/get — POISONING-003 scans the tool CALL RESULT itself. A compromised server can ship a tool whose annotations look benign (read-only hint on, no required arguments, boring name) and never show the payload in the catalog. The user sees a safe-looking "greeting" tool; the model sees a system prompt injection hidden in the greeting text. The check only runs when the audit was explicitly allowed to invoke tools (opt-in via --probe-tool-calls). Against third-party servers we skip this — calling a stranger's tool has side effects. Against our demo server and operator-authorised audits the check fires normally and finds the hidden characters and role markers that POISONING-001 and POISONING-002 would never see.

How to fix

Sanitise every tool output before returning it. Strip zero-width characters, bidi overrides, soft hyphens and tag characters. Treat upstream data (search results, scraped content, third-party APIs) as untrusted at the server layer and normalise with NFC + ASCII-print-filter before sending to the client. Never relay raw model-addressed text (System:, <SYSTEM>, "ignore previous instructions") through a tool result — if your business logic needs to emit such strings, tag them as a code span or quoted literal so the model does not read them as instructions.

See it live

Our demo MCP server trips POISONING-003 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/mcp --probe-tool-calls

The --probe-tool-calls flag is opt-in because tool invocation has side effects; safe against our demo, do not point at production servers you don't own.

Run 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.