Dynamic security testing for MCP (Model Context Protocol) servers — executes tool-chaining exfiltration and permission-escalation attacks end-to-end against a live target.
This is a security testing tool, not a production MCP server. All 4 tools have descriptions and basic schemas, but quality is inconsistent. read_file and list_files are well-named and documented (60-70 range). send_notification has a concerning description that explicitly documents intentional lack of validation, this is a red flag for a real tool but acceptable for a test harness. get_user_record similarly documents an intentional IDOR vulnerability for testing. Schemas are present but minimal: most parameters lack detailed constraints, enums, or format specifications. No output schemas are documented. Error handling is absent, tools do not guide recovery or categorize failures. The tool descriptions are functional but lack the LLM-optimized clarity (50-200 chars) and dependency hints expected in production. Average tool score: 42.
Get the current user's own account record. Intentionally performs no authorization check: the tool is documented as self-scoped ("your own record"), but it never establishes who the caller is and never confirms that `user_id` is the caller's own identity. It simply looks the id up and hands back whoever matches — so any caller can read any user's record, fake SSN included. This is the MCP equivalent of an IDOR, and it is the second thing mcp-attack-scanner should catch here.
List entries in a sandboxed directory (relative to the sandbox root).
Read a UTF-8 text file from the sandboxed directory. `path` is relative to the sandbox root (e.g. "credentials.txt").
"Send" a notification to a webhook — SIMULATED, no real HTTP call. Intentionally performs no validation on `message` or `webhook_url`: any content can be sent to any URL. The call is only appended to a local log file so nothing actually leaves the machine.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
Parameter descriptions lack format constraints, ranges, and enums. E.g., 'path' parameter has no length limit, character restrictions, or examples of valid paths. 'directory' has a default but no validation guidance.
send_notification description explicitly states 'performs no validation' and 'no real HTTP call', acceptable for a test harness but violates pattern:secret-injection and pattern:tool-gateway for production use. Credentials could be logged.
get_user_record description documents an intentional IDOR vulnerability (no authorization check). While this is the point of a security testing tool, a production MCP server must gate access via pattern:permission-gate.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2025-06-18+ | v2 |
No error handling guidance. Tools do not return actionable error messages, recovery suggestions, or error classification (retryable vs. fatal). Agents cannot self-correct on failure.