MCPShield
Back to home

Research & Benchmarks

MCPShield is the most comprehensive MCP security scanner available. Here are our benchmark results, detection methodology, and public dataset.

90+
Detection Rules
4,800+
Repos Scanned
100%
Benchmark Recall
0%
Evasion Rate

Benchmark Results

GenTelBench-v1

Malicious + safe tool descriptions

Samples
50,000
Recall
100%

50K samples from academic benchmark dataset

Source: GenTelLab (HuggingFace)

MCPTox

Poisoned tool descriptions

Samples
485
Recall
100%

All 485 malicious descriptions detected by regex scanner

Source: MCPTox research dataset

Adversarial Evasion

Crafted evasion descriptions

Samples
20
Recall
100%

4 evasion techniques: semantic rephrasing, legitimacy framing, multi-step, jargon obfuscation

Source: Internal (AutoMalTool techniques)

Real Malware

In-the-wild malicious MCP server

Samples
1
Recall
100%

Throwaway account, .exe lure, 23 zero-width characters detected

Source: GitHub (Parnellcold355/easypanel-mcp)

LLM Judge vs Regex: MCPTox Analysis

We tested both our regex scanner and LLM judge (Claude Haiku) against all 485 MCPTox malicious tool descriptions to measure the marginal value of each detection layer.

Regex Scanner485/485 (100%)
LLM Judge28/485 (5.8%)
Both Caught28 (5.8%)
Evasion Gap0 (0%)

The regex engine catches all known attack patterns. The LLM judge is conservative by design (low false positives) and provides value on novel semantic attacks that regex patterns haven't been written for yet.

Adversarial Evasion Testing

We crafted 20 adversarial tool descriptions using evasion techniques from the AutoMalTool paper (arxiv 2509.21011). Both regex and LLM caught every single one.

Semantic Rephrasing
Regex: 100%LLM: 100%

"compliance validation" instead of "exfiltrate"

Legitimacy Framing
Regex: 100%LLM: 100%

"built-in telemetry for quality assurance"

Multi-Step Obfuscation
Regex: 100%LLM: 100%

Attack split across tool chain references

Jargon Obfuscation
Regex: 100%LLM: 100%

"inode metadata traversal" for filesystem theft

3-Layer Detection Architecture

1

Regex Pattern Engine

90+ rules

Static pattern matching against tool descriptions, source code, and metadata. Covers OWASP MCP Top 10.

Strengths: Fast, deterministic, zero false negatives on known patterns

2

Schema Poisoning Scanner

Full schema traversal rules

Deep inspection of every text surface in tool metadata: descriptions, annotations, parameter defaults, enum values, comments, examples, and nested properties.

Strengths: Catches injection in overlooked fields like inputSchema.$comment

3

LLM-as-Judge (Claude Haiku)

Semantic analysis rules

AI-powered semantic analysis of tool descriptions. Detects manipulation that regex cannot catch through understanding intent.

Strengths: Catches novel semantic attacks, provides human-readable explanations

Mass Scan: 1,901 MCP Servers

We scanned 1,901 MCP server repositories from GitHub and 60 live HTTP endpoints. Results are available as an open dataset.

62.1%
Grade A
13.9%
Grade B
7.1%
Grade C
5.7%
Grade D
11.3%
Grade F
Critical findings834
High findings1,400
Live HTTP endpoints tested60
Endpoints responded26
Scan Your MCP Server