Skip to content
pwnkit pwnkit

pwnkit

Fully autonomous agentic pentesting framework. Attacks AI/LLM apps, web apps, package ecosystems, and source code. Blind PoC verification to minimize false positives.
Get Started View on GitHub

One command, zero config

Run pwnkit scan --target <url> and get a verified security report in minutes.

Blind verification

Every finding is independently re-exploited by a second agent that never sees the original reasoning. False positives are killed automatically.

Bring your own AI

Use your API key (OpenRouter, Anthropic, Azure OpenAI, OpenAI) or spawn Claude Code, Codex, or Gemini CLI with your existing subscription.

Full-spectrum pentesting

AI/LLM apps, web applications, package ecosystems, and source code repositories. Not just AI security — pwnkit covers traditional web vulnerabilities too. Cybench: 36/40 = 90.0% on the first scored full 40-challenge run (single-config gpt-5.4, single-shot). XBOW retained-artifact aggregate: 103/104 = 99.0% (only XBEN-030 unsolved in any mode), with the load-bearing gpt-5.4 cohort at 93/95 = 97.9% black-box at $5.20/flag (~$0.48/run). Historical mixed publication line documented separately on the benchmark page.

Explore the docs

Section titled “Explore the docs”

How it works

Section titled “How it works”
  • Architecture — the 5-stage pipeline, runtime adapters, and MCP integration
  • Agent Loop — how the autonomous loop thinks, calls tools, reflects, and stops
  • Finding Triage — the multi-layer pipeline between research and verify agents
  • Blind Verification — why findings are re-exploited before they count
  • Adversarial Evals — how pwnkit extends beyond classic pentesting into attack-driven evaluation of AI systems

Research and benchmark analysis

Section titled “Research and benchmark analysis”