Skip to content

v2026.4.0 — Git Scanning & Repo Poisoning Detection

Choose a tag to compare

View all tags
@rosschurchill rosschurchill released this 13 Mar 08:55
· 82 commits to main since this release
v2026.4.0
ab71c76

What's New

medusa scan --git <URL> — Scan Any Repo for AI Supply Chain Attacks

Scan any GitHub repository for weaponized AI editor configs, prompt injection payloads, and MCP tool poisoning with a single command:

pip install medusa-security
medusa scan --git user/repo

MEDUSA automatically detects 28+ AI editor config files that are known attack vectors across 15+ AI coding tools — Cursor, Cline, Copilot, Claude Code, Gemini CLI, Kiro, Codex CLI, Windsurf, Amazon Q, Roo Code, Tabnine, Continue.dev, and more.

Repo Poisoning Detection (18 new rules)

Detects real-world attacks:

  • Clinejection — GitHub Actions AI workflow abuse
  • CurXecute (CVE-2025-54135) — Cursor MCP auto-start RCE
  • IDEsaster (CVE-2025-64660) — VS Code YOLO mode injection
  • CamoLeak — Image URL data exfiltration
  • ToxicSkills — Malicious AI agent skills
  • AIShellJack — Shell built-in poisoning
  • Plus: fake conversation history, lock file backdoors, gist-piped-to-shell, secrecy directives, case-sensitivity bypass, JSON $schema exfiltration

MCP Advanced Attacks (11 new rules)

  • Full-schema tool poisoning, ATPA output poisoning, parameter name injection
  • ANSI escape terminal injection, sampling-based indirect injection
  • Cross-server manipulation, rug-pull detection, Flowise RCE (CVE-2025-26319)

Bug Fix: FP Filter Path Matching

Fixed false positive filter matching against full absolute paths — repos with "demo", "test", or "example" in their name now scan correctly.

Stats

  • 7,375 unique scanner rules across 43 categories
  • 28+ AI editor config file types detected
  • 37+ CVEs mapped to AI coding editor attacks
  • 57 new tests for the --git feature

Install / Upgrade

pip install --upgrade medusa-security
Assets 4
Loading