Skip to content

πŸ€– v2025.7.0.0 - AI Agent Security (50+ Detection Rules)

Choose a tag to compare

@rosschurchill rosschurchill released this 27 Nov 22:09
· 347 commits to main since this release

πŸ€– AI Agent Security Release

MEDUSA v2025.7.0.0 introduces industry-leading AI agent security scanning with 50+ detection rules for the agentic AI era.

πŸ†• New AI Security Scanners

Scanner Rules Detects
AI Context Scanner AIC001-030 Prompt injection, HITL bypass, semantic attacks, training poisoning
Agent Memory Scanner AIM001-010 Memory poisoning, insecure storage, checkpoint tampering
RAG Security Scanner AIR001-012 Knowledge base poisoning, embedding attacks, source confusion
MCP Config Scanner MCP001-013 Secrets, dangerous paths, untrusted sources, TLS issues
MCP Server Scanner MCP101-116 Tool poisoning, confused deputy, cross-server attacks

πŸ›‘οΈ Attack Coverage

Context & Input Attacks

  • Prompt injection patterns (ignore instructions, role manipulation)
  • Hidden instructions (HTML comments, Unicode tricks)
  • Obfuscation techniques (base64, rot13, invisible chars)

Tool & Action Attacks

  • Tool poisoning (CVE-2025-6514)
  • Tool name spoofing
  • Confused deputy patterns (CWE-441)
  • Cross-server attack vectors

Memory & State Attacks

  • Memory poisoning detection
  • Checkpoint tampering
  • Cross-session data exposure
  • Insecure vector store configs

RAG & Knowledge Attacks

  • Knowledge base injection patterns
  • Embedding pipeline security
  • Source validation bypass
  • Conflict resolution disabled

Advanced Attacks (NEW!)

  • HITL bypass techniques
  • Semantic manipulation (hidden meaning, loopholes)
  • Evaluation/ground truth poisoning
  • Training data/RLHF manipulation
  • Agent identity spoofing
  • Resource exhaustion attacks

πŸ“š Research-Backed

Based on extensive research from:

  • "Agentic Design Patterns" (500+ pages) via NotebookLM
  • MCP Security Research (CVE-2025-6514, Invariant Labs)
  • OWASP LLM Top 10

πŸš€ Quick Start

pip install --upgrade medusa-security

# Scan your AI config files
medusa scan .

πŸ“ Supported AI Files

.cursorrules           # Cursor AI
CLAUDE.md              # Claude Code
mcp.json               # MCP configs
*.mcp.ts / *.mcp.py    # MCP server code
rag.json               # RAG configs
memory.json            # Agent memory

Full Changelog: v2025.6.0.0...v2025.7.0.0

πŸ€– Generated with Claude Code