Pantheon-Security/medusa

What it is

Despite the catalog summary describing the Foofus/Medusa parallel login brute-forcer, the fetched page is a different tool that shares the name: MEDUSA, an "AI-first" security scanner distributed via pip. It ships 40,000+ detection patterns aimed at AI/ML code, LLM agents, MCP servers, and RAG pipelines, plus 200 CVE detections (Log4Shell, Spring4Shell, XZ Utils, LangChain RCE, MCP-Remote RCE, React2Shell). If you came here expecting the credential brute-forcer, this is not it — verify which tool you are actually pulling.

What it does

The two genuinely actionable capabilities, beyond the marketing feature list:

• Repo / supply-chain scanning — medusa scan --git <URL> scans a GitHub repo for AI supply-chain attacks: repo poisoning, prompt injection, and MCP tool poisoning, including weaponized AI-editor configs across 28+ file types (Cursor, Cline, Copilot, Claude Code, Gemini, Kiro).
• Leaked-secret scanning of AI/shell histories — medusa secrets scan greps AI-assistant chat logs (Claude Code, Cursor, Copilot, Zed, Gemini) and shell histories (bash, zsh, fish, psql, mysql, python REPL) for credentials that developers paste into assistants. It recognizes 21 issuer types (Anthropic, OpenAI, PyPI, GitHub PATs, AWS, GCP, Stripe, Slack). The threat model the page names: those conversations sit in plaintext under $HOME, so anyone with read access — or later malware — can harvest live production keys.

What to do

medusa secrets scan                       # everything (chat + shell)
medusa secrets scan --source ai-chats     # AI assistants only
medusa secrets scan --source shell        # ~/.bash_history, ~/.zsh_history, fish, psql, mysql
medusa secrets scan --path FILE           # explicit file (e.g. a ChatGPT export)
medusa secrets scan --reveal              # show real values (requires 'I UNDERSTAND')
medusa secrets purge                      # interactive [y/n/s/a/q] redaction

Purge writes a byte-identical backup before any change, keeps JSONL parseable, and tags each redaction with a unique per-run marker ([REDACTED-MEDUSA-...-<run-id>]). It is local-only with no telemetry.

Caveats

• This is vendor self-description on a project page; the pattern counts (40,000+, "10-40x faster"), CVE coverage, and false-positive hardening are unverified claims, not third-party results.
• The 40,000+ patterns are stated to be "harvested from 8,466 AI-security research papers" — coverage breadth, not tested detection quality.
• Name collision is the real hazard here: confirm you are installing the intended medusa, since the catalog entry points at the wrong tool.