AI writes the code.
pwnkit hacks it.

Open-source agentic harness for autonomous security research.

Agents that read code, craft attacks, and re-exploit to kill false positives · Built from 7 real CVEs in packages with 40M+ downloads

npx pwnkit-cli audit express or npx pwnkit-cli review ./my-repo

View on GitHub

Nightfang scanning a target and finding vulnerabilities

Terminal

$ npx pwnkit-cli scan --target https://demo.app/api/chat

pwnkit v0.1.0

Target: https://demo.app/api/chat

▸ Discover Found 3 endpoints, 2 MCP tools, system prompt extracted

▸ Attack Ran 47 test cases across 6 categories

▸ Verify Confirmed 4 of 7 findings (3 false positives eliminated)

▸ Report Written to ./pwnkit-report/

4 verified vulnerabilities found.

ID	Finding	Risk	Conf
NF-001	Direct prompt injection	HIGH	99%
NF-002	System prompt extraction	MED	95%
NF-003	MCP tool SSRF via fetch_url	HIGH	97%
NF-004	PII leak in chat context	HIGH	92%

One toolkit. Five attack surfaces.

From LLM endpoints to npm packages to git repos — pwnkit finds what scanners miss.

LLM Endpoints

ChatGPT, Claude, Llama APIs, custom chatbots

MCP Servers

Tool schemas, validation, auth, poisoning

npm Packages

Supply chain, malicious code, dependency risk

Source Code

Local repos, GitHub URLs, deep AI audit

Web Apps

AI copilots, RAG pipelines, agent APIs

Five commands. Full coverage.

Each command is purpose-built for a different attack surface. Zero config, instant results.

scan

Probe LLM endpoints & MCP servers

Discovers vulnerabilities in AI endpoints with 47+ test cases across prompt injection, jailbreaks, tool poisoning, data exfiltration, and more. Supports probe, deep, and MCP modes.

--mode probe|deep|mcp --depth quick|default|deep --agentic

npx pwnkit-cli scan --target <url>

audit

Audit npm packages for malicious code

Installs a package in a sandbox, runs semgrep static analysis plus AI-powered code review. Catches supply chain attacks, backdoors, and dependency vulnerabilities.

--version <ver> --depth quick|default|deep

npx pwnkit-cli audit <package>

review

Deep security audit of source code

Security-focused code review of local repos or GitHub URLs. Multiple AI runtimes analyze your entire codebase and output SARIF, Markdown, and JSON reports.

npx pwnkit-cli review <repo>

findings

Query and inspect verified findings

Filter findings by severity, category, and status. Inspect individual findings with full evidence chains and proof artifacts. Track the lifecycle from discovered to confirmed.

list --severity critical show NF-001

npx pwnkit-cli findings list

history

Browse past scan results

Query the local SQLite database for previous scans. See status, depth, findings count, and duration for every run. Track your security posture over time.

--limit 20 --db-path ./custom.db

npx pwnkit-cli history

One command, zero config

No YAML files. No Python environments. Just npx pwnkit-cli scan and you're running.

Zero false positives

Every finding is re-exploited with proof before it hits the report. No more triaging 200 "possible prompt injections."

$0.05 per CI scan

Quick scans in under a minute. Deep audits for $1. Cheaper than one hour of manual pentesting.

LLM agnostic

Works with any model — Claude, GPT, Ollama, Gemini, or your own fine-tune. Swap providers without changing a single config line.

How it compares

Independent. Open source. No vendor lock-in.

Feature	pwnkit	promptfoo (acquired by OpenAI)	garak	nuclei	Semgrep
Autonomous multi-agent	Agentic pipeline	—	—	—	—
Verification (no false positives)	Re-exploits	—	—	—	—
LLM endpoint scanning	✓	✓	✓	—	—
MCP server security	✓	—	—	—	—
npm package audit	✓	—	—	—	Rules
Source code review	AI-powered	—	—	—	Rules
Web/API scanning	✓	—	—	✓	—
AI attack coverage	30+ agentic	Partial	Partial	—	—
Zero config	npx	YAML	Python	Templates	Config
Independent	✓	Acquired	✓	✓	VC-backed
Open source	MIT	OpenAI-owned	OSS	MIT	LGPL

Drops into your CI/CD

Findings show up directly in GitHub's Security tab.

.github/workflows/pwnkit.yml

name: AI Security Scan
on: [push, pull_request]

jobs:
  pwnkit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run pwnkit
        uses: peaktwilight/pwnkit/action@v1
        with:
          target: $${{ secrets.STAGING_API_URL }}
      - name: Upload SARIF
        uses: github/codeql-action/upload-sarif@v3
        with:
          sarif_file: pwnkit-report/report.sarif

pwnkit reviews its own source code

Every 6 hours via GitHub Actions

pwnkit runs pwnkit review . on its own repository. The same agentic pipeline that found 7 CVEs — pointed at itself. If it finds something, you'll see it here.

View CI runs

Set it up on your repo in 2 minutes:

1. Add to your GitHub Actions workflow:

- run: npx pwnkit-cli review . --format json > pwnkit-report.json

2. Add the badge to your README:

[![pwnkit](https://pwnkit.com/badge/ORG/REPO)](https://pwnkit.com)

Built from real security research

pwnkit started as an internal framework. It found 7 CVEs in packages with 40M+ weekly downloads before I open-sourced it.

node-forge 32M/week mysql2 5M/week Uptime Kuma 86K stars LiquidJS CVE jsPDF 2 CVEs picomatch CVE

Full CVE writeups

Stop guessing.
Start proving.

Five commands. Real vulnerabilities. Proof of exploitability.


npx pwnkit-cli scan --target <url>

Star on GitHub

AI writes the code.pwnkit hacks it.