MCP Audit — Contract Validation & AI-Specific Threat Detection | 42Crunch
AI Security — Static Analysis
Catch what the AI model reads before it ever calls a tool.
MCP Audit is the static analysis engine for MCP Contracts. It runs automatically on every uploaded contract — no live connection, no flag to set — and produces a score, a grade, and a list of findings.
What's real today: content/agent-facing threat detection (prompt injection, tool poisoning, data exfiltration, tool shadowing) runs automatically, server-side, and produces genuine findings with severity and remediation. What's roadmap: structural contract checks — authentication scheme completeness, authorization coverage, tool integrity signing — are design-intent today, not yet scored. We say so below, section by section.
Two kinds of checks
Agent-facing vs. Backend
Every finding is classified by who it can hurt — the AI agent consuming your server, or the server's own security posture.
◈
Live today
Agent-facing — Content Analysis
Three engines analyze the free-text description of every tool and prompt in your contract, automatically, on every audit:
| PromptDefense | Checks for absence of defensive language across 12 attack-vector categories. |
| YARA | Matches known malicious signatures — prompt injection, credential harvesting, tool poisoning. |
| LLM opt-in | Semantic reasoning for novel threats signatures can't catch — requires an API key. |
Every content finding carries a remediation string with actionable guidance, shown directly in the Audit tab.
◉
Design intent — not yet scored
Backend — Structural Analysis
Eight contract sections have a designed rule set, not yet implemented. Today's score for these sections counts placeholder markers only — not real structural or security analysis.
server |
Naming, versioning conventions |
integrity |
Message signing, rug-pull protection |
authentication |
Scheme completeness, fail-closed defaults |
authorization |
Role coverage, wildcard-allow rejection |
capabilities |
Per-tool security block, risk level, confirmation requirements |
throttling |
Rate limiting, fail-closed on limit exceeded |
Compliance mapping
Every finding mapped to the frameworks you already report against.
Findings are tagged against OWASP, the EU AI Act, ISO 42001, NIST AI RMF, and CSA AICM automatically — so a security review doubles as compliance evidence without extra work.
5 / 5 frameworks mapped automatically
How it fits together
Audit is the first gate. Scan validates the live server.
Upload contract
|
v
MCP Audit ← content analysis runs here automatically
|
score + findings (Agent-facing / Backend)
|
v
SQG gate → CI/CD pass / block (single score threshold, default 75 / grade B)
|
v
MCP Scan (live protocol probing + fuzzing against the running server)
|
v
MCP Protect (contract enforced at runtime)
See MCP Scan for what runs against the live server, and MCP Protect for runtime enforcement.