Skip to content

🧠 v2025.8.0.0 - OWASP LLM 2025 + CVE-2025-6514 Detection

Choose a tag to compare

View all tags
@rosschurchill rosschurchill released this 28 Nov 08:38
· 347 commits to main since this release
v2025.8.0.0
84cf1b9

🧠 AI Security Powerhouse Release

MEDUSA v2025.8.0.0 doubles down on AI security with OWASP Top 10 for LLM 2025 compliance, CVE-2025-6514 detection, and comprehensive documentation.

🆕 What's New Since v2025.7

OWASP Top 10 for LLM Applications 2025

Updated OWASPLLMScanner to the November 2024 release of OWASP standards:

Category Status Description
LLM07 🆕 NEW System Prompt Leakage - credentials/secrets in prompts
LLM08 🆕 NEW Vector and Embedding Weaknesses - RAG security
LLM10 📝 Updated Unbounded Consumption (was "Model Theft") - DoS/DoW attacks
LLM01-06, LLM09 ✅ Enhanced Improved patterns for indirect injection, confused deputy

CVE-2025-6514 Detection (MCP117)

Critical OAuth command injection vulnerability in mcp-remote (CVSS 9.6):

// CRITICAL: MEDUSA now detects this attack pattern
const authUrl = authorization_endpoint + userInput;
open(authUrl);  // PowerShell subexpression injection on Windows

Affected: mcp-remote versions 0.0.5 - 0.1.15

Advanced Confused Deputy Detection (MCP118)

New patterns for privilege escalation via tool authorization:

  • Service account execution abuse
  • Auth header forwarding issues
  • Permission override patterns
  • Multi-hop authorization bypass

4 New GenAI Security Scanners (60 Total!)

Scanner Rules Focus
ModelAttackScanner 10 Adversarial inputs, model extraction, membership inference
LLMOpsScanner 10 Insecure model loading, checkpoint exposure, drift detection
VectorDBScanner 10 Unencrypted storage, tenant isolation, PII in embeddings
OWASPLLMScanner 10 Full OWASP Top 10 2025 compliance

📚 New Documentation

AI Security Guide (500+ lines)

Comprehensive documentation covering:

  • What Gets Scanned - File types, content triggers, detection flow
  • Practical Examples - LangChain, MCP servers, RAG pipelines, ML Ops
  • Scanning Workflows - CI/CD, pre-commit, security audits
  • Complete Rule Reference - All 150+ AI-specific rules

Example: Scanning a LangChain App

$ medusa scan my-chatbot/

CRITICAL (2)
  app.py:45          [LLM01] Prompt Injection: User input in f-string
  chains/agent.py:78 [LLM05] Improper Output: exec(llm_response)

HIGH (3)
  app.py:12          [LLM02] Hardcoded API key
  vectorstore/db.py:56 [VD002] Missing access controls
  ...

📊 By The Numbers

Metric v2025.7 v2025.8 Change
AI Scanners 12 16 +33%
AI Rules 50+ 150+ +200%
Total Scanners 56 60 +4
Documentation Basic Comprehensive +500 lines

🛡️ Standards Compliance

  • ✅ OWASP Top 10 for LLM 2025 (November 2024)
  • ✅ MITRE ATLAS (AI threat landscape)
  • ✅ CWE-441 (Confused Deputy)
  • ✅ CVE-2025-6514 (mcp-remote RCE)

🚀 Quick Start

pip install --upgrade medusa-security

# Scan your AI codebase
medusa scan .

# See AI-specific issues
medusa scan . --format html

📁 Supported AI Files

.py, .js, .ts, .tsx     # LLM application code
mcp.json                # MCP configurations
*.mcp.ts / *.mcp.py     # MCP server implementations
.cursorrules            # Cursor AI
CLAUDE.md               # Claude Code
rag.json / *.yaml       # RAG configurations

🔗 Resources

Full Changelog: v2025.7.0.0...v2025.8.0.0

🤖 Generated with Claude Code

Assets 2
Loading