mcp-taxonomy

Canonical classification taxonomy for the MCP security ecosystem.

Provides shared enums, types, and cross-project adapters so findings from palisade-scanner, MCPGuard, MCPwn, and agentgate can be correlated, compared, and displayed in a unified view (MCPscop).

Quick Start

from mcp_taxonomy import (
    AttackCategory, Severity, Confidence, DetectionMethod,
    palisade_finding_to_taxonomy,
    mcpguard_event_to_taxonomy,
    mcpwn_finding_to_taxonomy,
    agentgate_signal_to_taxonomy,
)

# Normalize findings from any tool into a common TaxonomyEvent
event = palisade_finding_to_taxonomy({
    "category": "jailbreak",
    "detector": "injection_patterns",
    "severity": "high",
    "confidence": 0.9,
    "title": "Jailbreak detected",
})

print(event.attack_category)  # AttackCategory.JAILBREAK
print(event.severity)         # Severity.HIGH
print(event.source)           # "palisade-scanner"

Taxonomy

Attack Categories (20)

Severity (5 levels)

critical (25) > high (10) > medium (3) > low (1) > info (0)

Detection Methods (22)

Hidden text, injection patterns, metadata, exfiltration, unicode advanced, stego markers, entropy, image stego, instruction classifier (palisade-scanner), prompt injection, jailbreak patterns, tool poisoning, resource prompt, stego detector, anomaly detector (MCPGuard), injection tester, prompt fuzzer, tool analysis, tool poisoning fuzzer, SSRF tester, SQLi tester, RCE blind tester, A2A scanner (MCPwn), known/suspicious UA, rate/honeypot (agentgate).

Adapters