AAK-MCP-CARD-001: Poisoned tool description in an MCP server card
- Severity
- CRITICAL
- Category
- MCP Server Card (
mcp-server-card) - Scanner
agent_audit_kit/scanners/mcp_server_card.py- Auto-fixable
- No.
- SARIF name
McpCardPoisonedTool
What it detects
A tool entry in a SEP-1649 server card (/.well-known/mcp/server-card.json) carries a description containing a tool-poisoning / imperative-injection payload — invisible Unicode, a prompt-injection directive, a cross-tool reference, or an encoded (base64/hex) blob. Because a client fetches and trusts the card before connecting, the injected instructions enter the model context ahead of any tool call. Detection reuses the AAK-POISON-001..006 detectors.
How to fix
Treat a server card's tool descriptions as untrusted content: strip invisible Unicode, reject imperative/injection phrasing and encoded payloads, and render descriptions as data (never as instructions). Pin the card to a signed, provenance-verified publisher.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-CARD-001
Limitations
Reads tools[].description in a committed card with the phrase lists of AAK-POISON-001..004, so the same wording gaps apply. Text in a tool's title, or in the parameter descriptions of its inputSchema, is not read.
References
- OWASP MCP Top 10: MCP03:2025 (Tool Poisoning)
- OWASP Agentic Top 10: ASI02 (Tool Misuse and Exploitation), ASI04 (Agentic Supply Chain Vulnerabilities)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-SKILL-01
All rules. This page is generated from the rule registry when the documentation is built; to change it, change the rule in agent_audit_kit/rules/builtin.py. To give the rule a hand-written write-up, add docs/rules/AAK-MCP-CARD-001.md: the page then shows that text, with the registry reference appended under it.