AAK-MCP-SAMPLING-001: MCP sampling capability declared without consent / elicitation guard
- Severity
- HIGH
- Category
- MCP Configuration (
mcp-config) - Scanner
agent_audit_kit/scanners/mcp_sampling_capability.py- Auto-fixable
- No.
- SARIF name
McpSamplingNoConsent
What it detects
An MCP server or client participates in the sampling capability (the server can request LLM completions through the client) without an accompanying elicitation/consent gate, human-approval flag, or documented risk acceptance. Sampling output must be treated as untrusted tool input — without a guard, a compromised or adversarial server can silently steer the client's LLM (and any tool calls it triggers).
How to fix
Require an elicitation/createMessage consent prompt before honoring sampling requests, and treat sampling output as untrusted tool input (re-run the same sanitization you apply to user input). If sampling is intentional, accept the risk in .agent-audit-kit.yml with accepts_sampling_risk: true and a non-empty justification:.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-SAMPLING-001
Limitations
Reads .py, .ts, .tsx, .js, .mjs, .cjs, .java, .kt and .rs files of a project whose root manifest names an MCP SDK in a form AAK-MCP-STATELESS-001 recognises, for a capabilities: { ... } literal that names sampling before its first closing brace, ServerCapabilities( with sampling= before its first ), SamplingCapability(, CreateMessageRequestSchema, sampling/createMessage, create_message(, .sampling.create( or RequestSampling. One consent marker in any of those files (elicit(, elicit_create(, elicit_input(, elicitation(, elicitation/create, ElicitRequestSchema, ElicitInput, requires_consent, require_consent, user_consent, human_in_the_loop, human-in-the-loop or humanintheloop in any case, confirmSampling, approveSampling, samplingConsent or human_approval) clears every one of those findings, whatever it guards. A source file over 512 KB, a config file over 1 MB or a manifest over 2 MB is not read.
References
- OWASP MCP Top 10: MCP06:2025 (Intent Flow Subversion)
- OWASP Agentic Top 10: ASI01 (Agent Goal Hijack), ASI09 (Human-Agent Trust Exploitation)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-AUTH-02
- CSA AI Controls Matrix (AICM): IAM-01, AIS-07
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-SAMPLING-001.md: the page then shows that text, with the registry reference appended under it.