Skip to contentAgent? Read agent.txt

Decoy Red Team

Autonomous adversarial testing for MCP servers. 53 attack patterns across 6 categories, mapped to OWASP Agentic Top 10.

Decoy Red Team connects to every MCP server on your machine, sends adversarial payloads to their tools, and reports what's exploitable. Not a scanner — an attacker. Every finding maps to OWASP Top 10 for Agentic Applications 2026.

What it tests

53 attack patterns across 6 categories:

CategoryWhat it tests
Input injectionSQL injection, command injection, path traversal, SSRF, template injection
Prompt injectionInstruction override, role hijack, indirect injection, encoding bypass, multi-turn
Credential exposure.env files, cloud credentials, SSH keys, git tokens, shell history
Protocol attacksMalformed JSON-RPC, capability escalation, replay attacks, method injection
Schema boundaryType coercion, null bytes, overflow, prototype pollution, NoSQL operators
Privilege escalationScope escape, undeclared access, dotfile enumeration, argument smuggling

Usage

Dry run by default — --live is required to actually execute anything.

# Show the attack plan without executing (safe, always)
npx decoy-redteam

# Execute attacks — prompts for confirmation
npx decoy-redteam --live

# Target a specific server
npx decoy-redteam --live --target=my-server

# Only test specific categories
npx decoy-redteam --live --category=input-injection,credential-exposure

# SARIF for CI
npx decoy-redteam --live --sarif > redteam.sarif

Safety

  • Dry-run by default. npx decoy-redteam plans but does not execute.
  • Live mode prompts for confirmation. No --yes bypass.
  • Safe by default. Live mode runs read-only and protocol attacks only. Destructive attacks (file writes, etc.) require --live --full with a second confirmation.

CI/CD

- uses: decoy-run/decoy-redteam@v1

Runs in safe mode against your MCP servers and uploads SARIF to the GitHub Security tab.

Upgrading to full coverage

Free mode tests the 53 deterministic patterns (Layer 1). Guard Pro unlocks ~198 AI-adaptive and encoding variants (Layer 2) — plans generated from your latest scan, adapted to each server's actual tool surface.

npx decoy-redteam --live --pro --token=$DECOY_TOKEN