🚀 MEDUSA v2025.9.0 - 6 New AI Security Scanners (70 Total)

Major Release: 6 New Research-Driven Security Scanners

MEDUSA now includes 70 independent security scanner implementations with 180+ AI-specific detection rules.

🆕 New Scanners

🔐 PostQuantumScanner (PQC001-PQC010)

Quantum-vulnerable cryptography detection - first open-source scanner for PQC readiness

• RSA, ECDSA, ECDH, Diffie-Hellman flagged as quantum-vulnerable
• Classical key sizes detected (RSA-2048, P-256 curves)
• Crypto-agility anti-patterns identified
• Recommends NIST FIPS 203/204/205 standards (ML-KEM, ML-DSA, SLH-DSA)

🎭 SteganographyScanner (STG001-STG010)

Hidden payloads in multimodal AI inputs

• Zero-width Unicode characters (\u200b, \u200c, \u200d, \ufeff)
• Control token injection ([INST], <|im_start|>, Human:, Assistant:)
• Homoglyph attacks (Cyrillic/Greek lookalikes)
• LSB steganography patterns
• Base64 payloads in prompts

🧠 ExcessiveAgencyScanner (EXA001-EXA010)

Over-permissioned AI agents - OWASP LLM Top 10 coverage

• Unrestricted tool access (tools: "*")
• Missing before_tool_callback validation
• Unbounded action loops
• Disabled human-in-the-loop controls
• Recursive agent calls without depth limits

🔌 PluginSecurityScanner (PLG001-PLG010)

Cross-Plugin Request Forgery (CPRF) detection

• Cross-plugin data access vulnerabilities
• Chat history exposure to plugins
• Plugin command injection
• Missing plugin authentication

⚙️ HyperparameterScanner (HPT001-HPT010)

ML training sabotage detection

• Extreme learning rates (>=1.0 or <=1e-7)
• Untrusted training configs from remote URLs
• Disabled regularization/early stopping
• Suspicious weight initialization

🐳 DockerMCPScanner (DKR001-DKR010)

Container security for MCP servers

• Root user detection
• Unpinned base images
• Exposed ports and volumes
• Missing security options

🔧 Enhanced Scanners

• OWASPLLMScanner - Added CVE-2024-5184, prompt obfuscation patterns
• ModelAttackScanner - Added CVE-2019-20634, CVE-2023-4969, GPU attacks
• MCPConfigScanner - Enhanced OAuth spec detection, new MCP patterns
• MCPServerScanner - Added PowerShell injection, more tool poisoning patterns
• AgentMemoryScanner - Memory poisoning, vector injection, cross-session attacks
• MultiAgentScanner - Prompt infection, LLM tagging, consensus bypass
• LLMOpsScanner - Ray/Shadow Ray CVEs, LoRA adapter security, GPU memory leaks

📊 Stats

Metric	Before	After
Total Scanners	64	70
AI Detection Rules	150+	180+
New Code	-	3,320 lines

📦 Installation

pip install --upgrade medusa-security

🔗 Links

• PyPI
• Documentation
• AI Security Guide