convez.ai
MCP Server
beejak

mcp-sentinel

beejak/mcp-sentinel

Ver no GitHub
BomPluginScore ยท de 0 a 100

๐Ÿ›ก๏ธ Enterprise-grade security scanner for Model Context Protocol (MCP) servers. Detects secrets, injection flaws, and vulnerabilities.

Instalaรงรฃo

$npx convez-ai beejak/mcp-sentinel

1 comando no terminal: detecta o formato do repositรณrio (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Mรฉtricas

Estrelas
2
Forks
0
Acompanhando
1
Saรบde
100%
รšltimo commit
hรก 4 meses
Versรฃo
v3.0.0
Issues abertas
0
Contribuidores
1

README

A traduรงรฃo para o portuguรชs deste README ainda estรก na fila.

MCP Sentinel

Version Python 3.9+ Tests License: MIT OWASP Code style: black

Static security scanner for Model Context Protocol (MCP) servers.

Finds hardcoded secrets, injection flaws, tool poisoning, rug pulls, and 10 more vulnerability classes โ€” before they reach production.

Quick Start ยท Detectors ยท Scan Report ยท CLI Reference ยท OWASP Coverage


See It In Action

$ mcp-sentinel scan ./my-mcp-server

โ•ญโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฎ
โ”‚ MCP Sentinel v0.5.0             โ”‚
โ”‚ Scanning: ./my-mcp-server       โ”‚
โ”‚ Engine:   static                โ”‚
โ•ฐโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฏ

 Scan Summary
โ•ญโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฎ
โ”‚ Files Scanned         โ”‚ 57/57        โ”‚
โ”‚ Duration              โ”‚ 0.09s        โ”‚
โ”‚ Total Vulnerabilities โ”‚ 86           โ”‚
โ•ฐโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฏ

 Vulnerabilities by Severity
โ•ญโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฎ
โ”‚ CRITICAL โ”‚    32 โ”‚
โ”‚ HIGH     โ”‚    45 โ”‚
โ”‚ MEDIUM   โ”‚     9 โ”‚
โ•ฐโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฏ

 OWASP Agentic AI Top 10 Coverage
โ•ญโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฎ
โ”‚ ASI ID  โ”‚ Category                                 โ”‚ Findings โ”‚ Max Severity โ”‚
โ”œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ค
โ”‚ ASI01   โ”‚ Prompt Injection                         โ”‚       52 โ”‚   CRITICAL   โ”‚
โ”‚ ASI02   โ”‚ Sensitive Data Exposure                  โ”‚        3 โ”‚   CRITICAL   โ”‚
โ”‚ ASI03   โ”‚ Supply Chain Vulnerabilities             โ”‚        1 โ”‚     HIGH     โ”‚
โ”‚ ASI04   โ”‚ Insecure Direct Tool Invocation          โ”‚       15 โ”‚   CRITICAL   โ”‚
โ”‚ ASI05   โ”‚ Improper Output Handling / SSRF          โ”‚        5 โ”‚   CRITICAL   โ”‚
โ”‚ ASI06   โ”‚ Excessive Agency                         โ”‚        1 โ”‚    MEDIUM    โ”‚
โ”‚ ASI08   โ”‚ Insecure Deserialization                 โ”‚        1 โ”‚   CRITICAL   โ”‚
โ”‚ ASI09   โ”‚ Improper Error Handling / Path Traversal โ”‚        8 โ”‚     HIGH     โ”‚
โ•ฐโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ•ฏ

 Risk Score: 82.4/100

 Found 86 vulnerabilities

Why MCP Sentinel

The MCP ecosystem is growing fast. Security tooling is not keeping up.

A scan of 2,614 public MCP server implementations found:

  • 82% had path traversal exposure
  • 67% had code injection surface
  • 30% were SSRF-vulnerable
  • 8,000+ were publicly accessible due to 0.0.0.0 binding
  • 30+ CVEs filed in Januaryโ€“February 2026 alone, including Anthropic's own reference implementations

MCP Sentinel catches the patterns behind these vulnerabilities โ€” in seconds, with no external dependencies and no data leaving your machine.


Quick Start

# Install
pip install mcp-sentinel

# Scan a directory
mcp-sentinel scan /path/to/your/mcp-server

# Scan and fail CI on critical/high only
mcp-sentinel scan . --severity critical --severity high --no-progress

# Export for GitHub Code Scanning
mcp-sentinel scan . --output sarif --json-file results.sarif

# Export OWASP compliance report
mcp-sentinel scan . --compliance-file compliance.json

Detection Coverage

14 detectors. Every finding maps to an OWASP Agentic AI Top 10 category and a CWE.

DetectorWhat It CatchesOWASPSeverity
SecretsDetectorAWS keys, OpenAI/Anthropic tokens, JWT, private keys, DB connection stringsASI02CRITICAL
CodeInjectionDetectoros.system(), subprocess(shell=True), eval(), exec(), SQL f-stringsASI04CRITICAL
PromptInjectionDetectorRole manipulation, system prompt exposure, jailbreak patterns, override directivesASI01HIGH
ToolPoisoningDetectorInvisible Unicode, sensitive path targeting, behavior overrides, cross-tool manipulationASI01CRITICAL
PathTraversalDetector../ sequences, zip slip, unvalidated file opens, unsafe os.path.join()ASI09HIGH
SSRFDetectorUnvalidated URLs, cloud metadata endpoints (169.254.169.254), open redirectsASI05CRITICAL
MissingAuthDetectorRoutes without auth decorators, sensitive paths, unauthenticated system toolsASI04HIGH
NetworkBindingDetector0.0.0.0 binding across Python, JS, Go, Java, and config filesASI06MEDIUM
ConfigSecurityDetectorDebug mode, open CORS, TLS disabled, weak secrets, exposed admin endpointsASI02HIGH
WeakCryptoDetectorMD5/SHA-1, ECB mode, insecure random, static IV, deprecated ciphersASI04HIGH
InsecureDeserializationDetectorpickle.loads(), yaml.load(), marshal, eval() as parser, PHP unserialize()ASI08CRITICAL
SupplyChainDetectorEncoded payloads, install-time exfiltration, silent BCC, typosquatted packagesASI03CRITICAL
MCPSamplingDetectorSampling misuse, sensitive data in LLM calls, prompt injection via samplingASI01HIGH
RugPullDetectorGlobal state mutation, first-call sentinel, time-based behavior evasionASI01CRITICAL

Scan Report: beejak / Vulnerable-MCP-Server

Tested against the world's first deliberately vulnerable MCP server โ€” 18 intentional vulnerabilities across 5 attack categories, based on real CVEs and novel MCP-specific attack patterns.

Summary

MetricValue
Files scanned57
Scan duration0.09s
Total findings86
Critical32
High45
Medium9
OWASP ASI categories covered8 / 10

OWASP Agentic AI Top 10 Coverage

ASICategoryFindingsSeverity
ASI01Prompt Injection52๐Ÿ”ด CRITICAL
ASI02Sensitive Data Exposure3๐Ÿ”ด CRITICAL
ASI03Supply Chain Vulnerabilities1๐ŸŸ  HIGH
ASI04Insecure Direct Tool Invocation15๐Ÿ”ด CRITICAL
ASI05Improper Output Handling / SSRF5๐Ÿ”ด CRITICAL
ASI06Excessive Agency1๐ŸŸก MEDIUM
ASI08Insecure Deserialization1๐Ÿ”ด CRITICAL
ASI09Improper Error Handling / Path Traversal8๐ŸŸ  HIGH

Selected Findings

Secrets (ASI02)

[CRITICAL] Hardcoded AWS Access Key
  config.py:47  ยท  SecretsDetector  ยท  CWE-798
  fake_aws_key: str = "AKIAFAKE1234567890AB"

[CRITICAL] Hardcoded AWS Secret Key
  config.py:48  ยท  SecretsDetector  ยท  CWE-798
  fake_aws_secret: str = "fakesecret/FAKE/abc123def456ghi789jkl012"

Code Injection (ASI04)

[CRITICAL] Command Injection via os.system()
  injection.py:135  ยท  CodeInjectionDetector  ยท  CWE-78
  Hint(2, "Create a malicious pickle object with os.system() in __reduce__")

[CRITICAL] Insecure Deserialization: pickle.loads() on Untrusted Data
  injection.py:295  ยท  InsecureDeserializationDetector  ยท  CWE-502
  obj = pickle.loads(decoded)  # noqa: S301 โ€” intentionally vulnerable

Tool Poisoning (ASI01)

[HIGH] Prompt Injection: System Prompt Exposure
  orchestrator.py:221  ยท  PromptInjectionDetector  ยท  CWE-94
  planning_task = f"""Break this task into sub-tasks for the agent system...

[HIGH] Path Traversal: Unsafe Path Joining
  orchestrator.py:103  ยท  PathTraversalDetector  ยท  CWE-22
  path = os.path.join(self.work_dir, tool_input["path"])

Rug Pull โ€” detected for the first time (ASI01)

[CRITICAL] Rug Pull: Time-Based Behavior Mutation
  rug_pull.py:153  ยท  RugPullDetector  ยท  CWE-913
  elapsed = now - _compliance_first_call
  โ†’ Tool returns safe response during scanner window, exfiltrates after delay

[HIGH] Rug Pull: Global State Mutation
  rug_pull.py:112  ยท  RugPullDetector  ยท  CWE-913
  global _analyse_call_count
  โ†’ Call #1 returns clean result; call #2+ exfiltrates to attacker-controlled server

[MEDIUM] Rug Pull: First-Call Sentinel Pattern
  rug_pull.py:150  ยท  RugPullDetector  ยท  CWE-913
  if _compliance_first_call is None:

Infrastructure (ASI06)

[MEDIUM] Network Binding: Server Exposed on All Interfaces (0.0.0.0)
  docker-compose.yml:20  ยท  NetworkBindingDetector  ยท  CWE-605
  MCP_HOST: "0.0.0.0"

Full JSON report, SARIF file, and OWASP compliance report are available in reports/. The Vulnerable-MCP-Server README includes a full breakdown with remediation guidance: beejak/Vulnerable-MCP-Server


Output Formats

Terminal (default)

Colour-coded table with severity levels, file locations, code snippets, and remediation steps. Best for interactive review.

mcp-sentinel scan /path/to/server

JSON

Structured output with all findings, metadata, CVSS scores, MITRE ATT&CK IDs, and OWASP mappings. Pipe into any downstream tool.

mcp-sentinel scan . --output json --json-file results.json
{
  "type": "secret_exposure",
  "title": "Hardcoded AWS Access Key",
  "severity": "critical",
  "confidence": "high",
  "file_path": "config.py",
  "line_number": 47,
  "code_snippet": "aws_key: str = \"AKIA...\"",
  "cwe_id": "CWE-798",
  "cvss_score": 9.1,
  "owasp_asi_id": "ASI02",
  "owasp_asi_name": "Sensitive Data Exposure",
  "remediation": "Revoke this key immediately. Use environment variables or a secrets manager.",
  "fixed_code": "aws_key: str = os.getenv(\"AWS_ACCESS_KEY_ID\")",
  "mitre_attack_ids": ["T1552.001"],
  "detector": "SecretsDetector",
  "engine": "static"
}

SARIF 2.1.0

Compatible with GitHub Code Scanning, GitLab SAST, and Azure DevOps. Upload directly to the GitHub Security tab.

mcp-sentinel scan . --output sarif --json-file results.sarif

GitHub Actions:

- name: MCP Sentinel Scan
  run: |
    pip install mcp-sentinel
    mcp-sentinel scan . --output sarif --json-file results.sarif

- name: Upload to GitHub Security
  uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: results.sarif

OWASP Compliance Report

Exports a structured JSON report mapping every finding to its ASI01โ€“ASI10 category, with severity breakdown and coverage gaps.

mcp-sentinel scan . --compliance-file compliance.json
{
  "framework": "OWASP Agentic AI Top 10 2026",
  "total_findings": 86,
  "categories": {
    "ASI01": {
      "name": "Prompt Injection",
      "finding_count": 52,
      "max_severity": "critical",
      "severity_breakdown": { "critical": 25, "high": 23, "medium": 4 }
    },
    "ASI02": {
      "name": "Sensitive Data Exposure",
      "finding_count": 3,
      "max_severity": "critical"
    }
  }
}

CLI Reference

Usage: mcp-sentinel [OPTIONS] COMMAND [ARGS]...

Options:
  --log-level [DEBUG|INFO|WARN|ERROR|FATAL]
                        Logging verbosity. Use DEBUG to trace pattern matches.
                        [default: INFO]
  --log-file PATH       Write logs to a file in addition to stderr.
  --version             Show version and exit.
  --help                Show this message and exit.

Commands:
  scan    Scan a directory or file for security vulnerabilities.

mcp-sentinel scan

Usage: mcp-sentinel scan [TARGET] [OPTIONS]

  Scan a directory or file for security vulnerabilities.

  TARGET is the path to scan โ€” a directory or a single file.
  If omitted, mcp-sentinel will prompt you interactively.

  Runs 14 pattern-based detectors covering: hardcoded secrets, code
  injection, prompt injection, tool poisoning, path traversal, config
  security, SSRF, network binding, missing auth, supply chain attacks,
  weak cryptography, insecure deserialization, MCP sampling misuse,
  and rug pull / timed evasion attacks.

Arguments:
  TARGET    Path to scan (file or directory). Prompts if omitted.

Options:
  -o, --output [terminal|json|sarif]
                        Output format.
                          terminal  Colour-coded table โ€” best for interactive use.
                          json      Structured findings โ€” use for scripting.
                          sarif     SARIF 2.1.0 โ€” use for GitHub Code Scanning.
                        [default: terminal]

  --severity [critical|high|medium|low|info]
                        Only show findings at or matching the given severity.
                        Repeatable: --severity critical --severity high
                        Omit to show all severities.

  --json-file PATH      Output file for json or sarif formats.
                        Required when --output sarif (for GitHub upload).
                        If omitted with --output json, prints to stdout.

  --no-progress         Suppress the animated progress bar.
                        Use in CI environments to keep logs clean.

  --compliance-file PATH
                        Write an OWASP Agentic AI Top 10 compliance report
                        (JSON) to this file. Lists all ASI01โ€“ASI10 categories
                        with finding counts, severity breakdown, and gaps.

  --help                Show this message and exit.

Common workflows

# Interactive review โ€” default terminal output
mcp-sentinel scan /path/to/mcp-server

# CI โ€” fail on critical/high, suppress noise
mcp-sentinel scan . --severity critical --severity high --no-progress

# GitHub Code Scanning integration
mcp-sentinel scan . --output sarif --json-file results.sarif

# Export all findings as JSON
mcp-sentinel scan . --output json --json-file results.json

# Scan a single file with debug logging
mcp-sentinel --log-level debug scan server.py

# Keep audit log while reviewing interactively
mcp-sentinel --log-file audit.log scan .

# OWASP Agentic AI Top 10 compliance report
mcp-sentinel scan . --compliance-file compliance.json

# All outputs at once
mcp-sentinel scan . \
  --output json --json-file results.json \
  --compliance-file compliance.json \
  --no-progress

OWASP Agentic AI Top 10 Coverage

Every finding is annotated with its ASI category. The table below shows which detectors map to which categories.

ASICategoryDetectors
ASI01Prompt InjectionPromptInjectionDetector, ToolPoisoningDetector, RugPullDetector, MCPSamplingDetector
ASI02Sensitive Data ExposureSecretsDetector, ConfigSecurityDetector
ASI03Supply Chain VulnerabilitiesSupplyChainDetector
ASI04Insecure Direct Tool InvocationCodeInjectionDetector, MissingAuthDetector, WeakCryptoDetector
ASI05Improper Output Handling / SSRFSSRFDetector
ASI06Excessive AgencyNetworkBindingDetector
ASI07Insecure Plugin DesignToolPoisoningDetector, ConfigSecurityDetector
ASI08Insecure DeserializationInsecureDeserializationDetector
ASI09Improper Error Handling / Path TraversalPathTraversalDetector

Architecture

mcp-sentinel scan .
        โ”‚
        โ–ผ
MultiEngineScanner
        โ”‚
        โ–ผ
StaticAnalysisEngine
        โ”‚
        โ”œโ”€โ”€ SecretsDetector              AWS/OpenAI/GitHub keys, JWTs, PEM blocks
        โ”œโ”€โ”€ CodeInjectionDetector        os.system, subprocess(shell=True), eval, exec, SQL
        โ”œโ”€โ”€ PromptInjectionDetector      role manipulation, system prompt exposure, jailbreaks
        โ”œโ”€โ”€ ToolPoisoningDetector        invisible Unicode, path targeting, cross-tool manipulation
        โ”œโ”€โ”€ PathTraversalDetector        ../ sequences, zip slip, unsafe joins โ€” with taint analysis
        โ”œโ”€โ”€ SSRFDetector                 unvalidated URLs, cloud metadata endpoints
        โ”œโ”€โ”€ MissingAuthDetector          routes without auth decorators, sensitive paths
        โ”œโ”€โ”€ NetworkBindingDetector       0.0.0.0 binding across Python / JS / Go / Java / config
        โ”œโ”€โ”€ ConfigSecurityDetector       debug mode, open CORS, TLS off, weak secrets
        โ”œโ”€โ”€ WeakCryptoDetector           MD5/SHA-1, ECB, insecure random, static IV
        โ”œโ”€โ”€ InsecureDeserializationDetector  pickle, yaml.load, marshal, eval-as-parser
        โ”œโ”€โ”€ SupplyChainDetector          encoded payloads, install-time exec, BCC injection
        โ”œโ”€โ”€ MCPSamplingDetector          sampling misuse, LLM call injection
        โ””โ”€โ”€ RugPullDetector              global state mutation, first-call sentinel, timed evasion
        โ”‚
        โ–ผ
SeverityCalibrator          elevates severity based on MCP server context
        โ”‚
        โ–ผ
Findings โ†’ Deduplication โ†’ OWASP annotation โ†’ Output formatter

Design principles:

  • No external binaries โ€” pure Python stdlib + pydantic, click, rich
  • No network calls โ€” all analysis runs locally, nothing leaves your machine
  • Async scanning โ€” concurrent file I/O with configurable worker pool
  • MD5-based file cache โ€” unchanged files skipped on successive scans
  • Test-file awareness โ€” tests/, fixtures/, spec/ directories auto-suppressed to eliminate assertion false positives
  • Python AST โ€” multi-line subprocess(shell=True) detected via stdlib AST, not just regex

Supported Languages

LanguageExtensions
Python.py
JavaScript / TypeScript.js, .jsx, .ts, .tsx
Go.go
Java.java
Configuration.yaml, .yml, .json, .env, .toml
Shell.sh, .bash

Development

# Clone and install with dev dependencies
git clone https://github.com/beejak/mcp-sentinel.git
cd mcp-sentinel
pip install -e ".[dev]"

# Run the full test suite (619 tests, ~8s)
python -m pytest tests/ -v

# Run a specific detector's tests
python -m pytest tests/unit/test_tool_poisoning.py -v

# Lint
ruff check src/
black --check src/

# Type check
mypy src/

# Run against the deliberately vulnerable MCP server
git clone https://github.com/beejak/Vulnerable-MCP-Server ../Vulnerable-MCP-Server
mcp-sentinel scan ../Vulnerable-MCP-Server

Test suite: 619 passing, 4 xfail (multi-line taint flows that require semantic analysis โ€” tracked in tests/unit/test_path_traversal.py)

Adding a detector

1. Create  src/mcp_sentinel/detectors/your_detector.py   โ€” extend BaseDetector
2. Add     src/mcp_sentinel/detectors/__init__.py         โ€” import + __all__
3. Register src/mcp_sentinel/engines/static/static_engine.py  โ€” add to _get_default_detectors()
4. Write   tests/unit/test_your_detector.py               โ€” at least 5 assertions

Threat Model & Limitations

MCP Sentinel is a static analysis tool. It finds patterns in source code โ€” it does not execute code or observe runtime behavior.

What it catches:

ClassDetector
Hardcoded secretsSecretsDetector
Code / command injectionCodeInjectionDetector
Prompt injectionPromptInjectionDetector
Tool poisoning (all schema fields)ToolPoisoningDetector
Path traversalPathTraversalDetector
SSRFSSRFDetector
Missing authenticationMissingAuthDetector
0.0.0.0 network bindingNetworkBindingDetector
Insecure configurationConfigSecurityDetector
Weak / broken cryptographyWeakCryptoDetector
Insecure deserializationInsecureDeserializationDetector
Supply chain attacksSupplyChainDetector
MCP sampling misuseMCPSamplingDetector
Rug pull / timed evasionRugPullDetector

What it does not catch:

  • Multi-line taint flows (x = request.args.get("f") on line 1, open(x) on line 50) โ€” requires semantic analysis
  • Runtime rug pulls using external state (database, remote config) rather than module-level variables
  • Logic flaws and business logic vulnerabilities
  • Deep transitive dependency analysis (direct manifest files and code patterns only)

What's New

v0.1v0.2v0.3v0.4v0.5
Detectors69101214
Tests248334409525619
OWASP ASI mappingโ€”โ€”โ€”โ€”โœ…
Rug pull detectionโ€”โ€”โ€”โ€”โœ…
Test-file FP suppressionโ€”โ€”โ€”โ€”โœ…

v0.5.0 โ€” RugPullDetector (global state mutation, first-call sentinel, time-based evasion), OWASP Agentic AI Top 10 mapping on all findings, 21% false positive reduction via test-file awareness, MCP sampling audit.

v0.4.0 โ€” WeakCryptoDetector (MD5/SHA-1, ECB, insecure random, deprecated ciphers), InsecureDeserializationDetector (pickle, yaml.load, marshal, PHP unserialize, Node.js VM escape).

v0.3.0 โ€” SupplyChainDetector (encoded payloads, install-time exfiltration, BCC injection, typosquatted packages).

v0.2.0 โ€” SSRFDetector, NetworkBindingDetector, MissingAuthDetector, full-schema tool poisoning.


References


Security

To report a vulnerability in MCP Sentinel itself, see SECURITY.md.

Docker Deployment Hardening (2026-04-20)

docker-compose.yml no longer ships with hardcoded default credentials. All sensitive values now use Docker's :? syntax โ€” compose will exit with an error rather than silently falling back to a weak default if any required variable is unset:

DB_PASSWORD:?DB_PASSWORD is required
REDIS_PASSWORD:?REDIS_PASSWORD is required
SECRET_KEY:?SECRET_KEY is required
MINIO_ROOT_USER:?MINIO_ROOT_USER is required
MINIO_ROOT_PASSWORD:?MINIO_ROOT_PASSWORD is required

A .env.example file has been added to the repo root. Before running the stack:

cp .env.example .env
# Fill in every required value in .env, then:
docker-compose up -d

The .env file is gitignored and must never be committed.


License

MIT โ€” see LICENSE.

Ver o original no GitHub

Bom para

Relacionados