pwnkit-cli

pwnkit

Let autonomous AI agents hack you so the real ones can't.
Fully autonomous agentic pentesting framework. Blind PoC verification to minimize false positives.

Docs · Website · Blog · Benchmark

Autonomous AI agents that pentest web apps, LLM endpoints, npm packages, and source code. The agent gets a bash tool and acts like a real pentester — writing curl commands, Python exploit scripts, and chaining vulnerabilities. Every finding is independently re-exploited by a blind verify agent to kill false positives.

npx pwnkit-cli

Quick Start

# Pentest a web app
npx pwnkit-cli scan --target https://example.com --mode web

# Scan an LLM endpoint
npx pwnkit-cli scan --target https://your-app.com/api/chat

# Audit an npm package
npx pwnkit-cli audit lodash

# Review source code
npx pwnkit-cli review ./my-app

# Auto-detect — just give it a target
npx pwnkit-cli https://example.com
npx pwnkit-cli express
npx pwnkit-cli ./my-repo

See the documentation for configuration, runtime modes, and CI/CD setup.

How It Works

The agent gets 3 tools: bash, save_finding, done. It runs curl, writes Python scripts, chains exploits — the same way a human pentester works. No templates, no static rules.

  Research Agent              Blind Verify Agent           Report
  discover + attack + PoC --> gets ONLY PoC + path    --> SARIF / JSON / MD
                              no reasoning, no bias       only confirmed findings
                              can't reproduce? killed

The blind verification is the differentiator. The verify agent can't be biased by the research agent's reasoning.

Benchmark

XBOW (traditional web vulnerabilities)

Tested against the XBOW benchmark — 104 Docker CTF challenges covering SQLi, IDOR, SSTI, SSRF, file upload, deserialization, auth bypass, and more.

35 flags extracted across IDOR, SQLi, blind SQLi, SSTI, RCE, SSRF, LFI, XXE, file upload, deserialization, auth bypass, business logic, and cookie manipulation.

Tool	Score	Approach
Shannon	96.15%	White-box, source-aware
KinoSec	92.3%	Black-box, proprietary
XBOW	85%	Purpose-built
Cyber-AutoAgent	84.62%	Open-source, meta-agent
pwnkit	testing	Open-source, shell-first

AI/LLM security

10 custom challenges covering prompt injection, jailbreaks, system prompt extraction, PII leakage, encoding bypass, multi-turn escalation, MCP SSRF.

100% (10/10) — all flags extracted, zero false positives.

pnpm bench --agentic    # AI/LLM benchmark

See benchmark details.

GitHub Action

- uses: peaktwilight/pwnkit@main
  with:
    mode: review
    path: .
    format: sarif
  env:
    OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}

Built By

Created by a security researcher with 7 published CVEs. pwnkit exists because modern attack surfaces require agents that adapt, not static rules that don't.

Contributing

git clone https://github.com/peaktwilight/pwnkit.git
cd pwnkit && pnpm install && pnpm test

See CONTRIBUTING.md for guidelines.

Built by Peak Twilight -- also building FoxGuard, vibecheck, unfuck, whatdiditdo

License

Apache 2.0