HackMyAgent

OpenA2A: CLI · HackMyAgent · Secretless · AIM · Browser Guard · DVAA

Website · Demos · Discord

Security scanner for AI agents. It finds the shadow AI on a machine, scans agent projects and packages for credentials, risky MCP configs, prompt injection and weak governance, and tests agents with attack payloads.

Quick start

Scan the project you are in:

npx hackmyagent secure

  support-agent  v1.0.0 · library · 5 files analyzed
  [...]
  ── Observations ────────────────────────────────────────────
  Surfaces    library · 5 semantic artifacts · all 5 artifacts reached 0-6 of 7 analyzer families
  Checks      261 static declared · 63 of 63 check groups ran · 5 semantic (NanoMind AST, 0-6 of 7 analyzer families) · 5 files read by static checks
  [...]
  ── Findings ────────────────────────────────────────────────
  [...]
  │ CRITICAL  Overprivileged MCP server scope
  │ .mcp.json:5
  │ Overprivileged filesystem access allows the agent (or an attacker via prompt injection) to read sensitive files like SSH keys, credentials, and system configs.
  │ Verify: sed -n '5p' /home/dev/workspace/support-agent/.mcp.json
  │ Fix: Scope "filesystem" to the project directory: replace "/" with "./" or a specific subdirectory.
  [...]
  Path forward: 53 -> 92 by fixing 1 critical + 3 high

No config file or flag is needed. [...] marks omitted lines. The full output, the sample workspace and the build that produced it are in docs/quick-start-capture.txt. The run exits 1 because a critical or high finding fired.

What do you want to check?

• My own project, before I ship it: npx hackmyagent secure. Next: npx hackmyagent secure --fix.
• A package, repository or skill, before I install it: npx hackmyagent check . Next, for a local copy: npx hackmyagent secure --fix.
• A running agent endpoint you own: npx hackmyagent attack . Next, with every payload: npx hackmyagent attack --intensity aggressive.
• One text, such as a pull request body: npx hackmyagent scan-text . Next, in a pipeline: npx hackmyagent scan-text - --as pr-body --json.
• Every agent project under this folder: npx hackmyagent detect. Next, for one project: npx hackmyagent secure .
• More. A SOUL.md governance file: npx hackmyagent scan-soul. Attack payloads to run yourself against a skill, MCP config or SOUL.md: npx hackmyagent red-team . MCP endpoints a host you own exposes: npx hackmyagent scan .

What it finds

• Shadow AI: the AI assistants, MCP servers and agent projects on a machine (detect). Example: 1 project MCP server with sensitive access.
• Credentials and risky configuration: 261 static checks, from MCP configs to supply chain and sandbox escape. Example: CRITICAL Overprivileged MCP server scope.
• Undeclared capabilities and scanner evasion: 29 semantic checks over skills, prompts and configs. Example: CRITICAL Data Exfiltration Surface.
• Weak governance: SOUL.md graded against a 9-domain model (scan-soul). Example: Missing SOUL-IH-003 Role-play refusal.
• Text that gives orders: pull-request bodies, issues and comments (scan-text). Example: TEXT-001 HIGH Instruction-override payload.
• Attacks an agent falls for: 164 adversarial payloads (attack). Example: Prompt Injection: 2/2 successful against a deliberately vulnerable agent.

Usage

Fix what a scan finds

hackmyagent secure --fix --dry-run   # preview the fixes
hackmyagent secure --fix             # apply them
hackmyagent rollback                 # revert them

Backups go to a store outside the project. Which checks it fixes, and where the backups live: docs/SECURITY_CHECKS.md.

• hackmyagent secure: scores a sample agent project, then applies the fixes it can: text version, provenance.

Find the shadow AI on a machine

npx hackmyagent detect

  workspace  shadow ai audit · laptop · 2 agents · 3 agent projects
  [...]
  ── Shadow AI agents (3) ────────────────────────────────────
  project              identified by  mcp servers     governance   cred      verdict
  support-agent        Claude Code    2 mcp critical  gov   0/100  cred no   CRITICAL
  [...]

One row per agent project, worst first. Full output: docs/quick-start-capture.txt.

Gate a pull request in CI

name: Agent Security
on: [push, pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: '20' }
      - run: npx hackmyagent secure --json > security-report.json
      - run: npx hackmyagent secure -b oasb-1 --fail-below 70

secure exits 1 on a critical or high finding or a score below --fail-below, and 2 when it did not examine everything it found. Exit codes: docs/commands.md. More recipes: docs/ci.md.

Use it from an AI assistant

hackmyagent init-mcp --root /absolute/path/to/your/project

That adds the server to Claude Code, Cursor or VS Code. Then ask the assistant: “Run a deep security scan on this project.” The server reads only inside its roots, and no MCP tool writes to your files.

Install

npm install -g hackmyagent
brew install opena2a-org/tap/hackmyagent

Installing runs the install script of onnxruntime-node, which on Linux x64 downloads an execution-provider package from nuget.org. Node.js versions, skipping that download, and verifying a release: docs/install.md.

Telemetry

Usage telemetry is on by default, and hackmyagent --version says so on its second line: Telemetry: on (opt-out: OPENA2A_TELEMETRY=off ...).

• Sent: one event per command, to the OpenA2A Registry.
• Fields, and only these: tool, version, install ID, event type, command name, success, duration, platform, Node major version.
• Failed command: an error code is added to the command name.
• Install ID: derived from a SHA-256 hash of the machine identifier, or of the hostname where none can be read.
• Not sent: the raw identifier, file paths, scanned content, arguments beyond the command name, environment variables.
• Off under CI: CI, GITHUB_ACTIONS and the other CI services’ variables; --version then prints Telemetry: off.
• Off under test runners: Vitest, Jest, node --test, NODE_ENV=test; --version then prints Telemetry: off.
• On under CI or a test runner: only when OPENA2A_TELEMETRY=on is set for the run.
• Policy: opena2a.org/telemetry.

OPENA2A_TELEMETRY=off hackmyagent secure          # off for one run
hackmyagent telemetry off                         # off on this machine; `telemetry status` shows the state
OPENA2A_TELEMETRY_DEBUG=print hackmyagent secure  # print each event on stderr before it is sent

Sharing scan findings with the Registry is a separate opt-in (--contribute, or contribute.enabled in ~/.opena2a/config.json; --no-contribute or --ci turns it off for a run). hackmyagent telemetry off covers usage telemetry only.

Documentation

• Commands: the main commands and their flags, the MCP server, use through opena2a-cli, and exit codes.
• CI/CD: non-interactive runs, pull-request gates, credentials in git history, and scanning a pipeline’s text.
• How it works: discovery, static checks, the semantic layer, scoring, the simulation and the self-integrity check.
• Install: Node.js versions, Homebrew, building from source, the install-time download, verifying provenance.
• Security checks: check categories with representative checks, what each count includes, and what --fix changes.
• Use cases: scanning an agent, red-teaming an MCP server, securing OpenClaw and wiring a CI pipeline.
• Programmatic API: TypeScript entry points for the scanner, runtime protection and the NanoMind semantic compiler; plugin authoring.
• Contributing: outside pull requests are welcome. Report security issues through SECURITY.md, not a public issue.

Part of the OpenA2A security platform. Related: opena2a-cli, the unified CLI that runs HackMyAgent behind opena2a review; aicomply, which classifies PII, credentials and regulated data in agent I/O at runtime: HackMyAgent scans the code, aicomply guards the live stream. More in the OpenA2A docs and research.

License

Apache-2.0. See LICENSE.