Open Kritt: AI Agents That Actually Find Real Vulnerabilities

By Prahlad Menon 5 min read

Most AI security scanners produce noise. Point a model at an entire repo and ask for vulnerabilities, and you get hallucinated CVEs and false positives that waste more time than they save.

Open Kritt takes a different approach: break security research into small, focused tasks, run them across AI agents in parallel, and combine outputs into findings you can actually validate.

The credibility backing this tool: the Kritt team has earned $1.5 million in bug bounty payouts under the researcher name Blockian. Open Kritt is the open-source version of their internal tooling.

GitHub: Kritt-ai/open-kritt

How It Works

  1. Build workflows — Chain focused prompts into reusable security research playbooks
  2. Run scans — Analyze repos (remote or local) and dependencies with Codex or Claude Code
  3. Verify findings — Post-scripts validate issues, build PoCs, and generate reports
  4. Prioritize results — Custom severity rankers, consistent finding schema, automatic deduplication

The key insight: instead of asking “find all vulnerabilities,” you define specific security checks as workflow steps. Each step gets a focused prompt, runs against the codebase, and feeds into the next step. The agents work in parallel where possible.

Why This Matters

Traditional static analysis tools find syntax-level issues but miss logic bugs. LLMs can reason about code semantics but hallucinate when given too much context.

Open Kritt sits in the middle:

  • Focused prompts keep context windows manageable
  • Workflow chaining builds up context incrementally
  • Post-scripts verify findings before reporting them
  • Deduplication prevents the same issue from appearing 50 times

This is how professional security researchers actually work—they don’t scan the whole repo at once. They focus on high-risk areas, validate findings manually, and build exploits to confirm severity.

Model Support

Bring your own access:

  • Codex — Use your existing Codex login
  • Claude Code — Via Anthropic API
  • OpenAI — GPT models via API
  • OpenRouter — Access multiple providers

No vendor lock-in. You control the model, the prompts, and the infrastructure.

Self-Hosted

Open Kritt runs locally via Docker:

git clone https://github.com/Kritt-ai/open-kritt
cd open-kritt
./kritt setup
./kritt start

Then open http://localhost:5173. The setup wizard walks you through model access configuration.

Security Warning

Agents run as root in disposable job containers with internet access—they need this to install tools, compile code, and build PoCs. Run on a dedicated Docker host or VM. Don’t scan untrusted code on your main machine. Read the threat model before deploying.

Who This Is For

  • Security researchers doing bug bounty or audit work
  • Security-minded developers who want deeper analysis than SAST tools provide
  • Teams building security automation into CI/CD
  • Anyone who wants AI-assisted security research without SaaS lock-in

GitHub: Kritt-ai/open-kritt
License: AGPL-3.0
Requirements: Docker, Node.js 20+, model API access

Frequently Asked Questions

Is Open Kritt free?

Yes. Open Kritt is open source under the AGPL-3.0 license. You self-host it and bring your own model access (Codex login, Anthropic API key, OpenAI API key, or OpenRouter). You pay your model provider directly—there’s no Kritt fee.

How is this different from running Claude on my code?

Pointing a model at an entire repository and asking for vulnerabilities produces mostly noise. Open Kritt breaks the work into focused workflow steps, each with a specific prompt and scope. It also validates findings with post-scripts, deduplicates results, and applies severity ranking. This is the difference between “ask once and hope” and “systematic security research.”

What types of vulnerabilities can it find?

Open Kritt can find logic bugs, access control issues, injection flaws, dependency vulnerabilities, and other semantic issues that traditional SAST tools miss. The specific coverage depends on your workflows—you define the prompts and research steps. The Kritt team’s $1.5M in bounties came primarily from smart contract and web3 security.

Is it safe to run on my codebase?

Open Kritt agents run as root in containers with internet access, so they can install dependencies and build PoCs. This is necessary for security research but means you should run it on a dedicated Docker host or VM, not your development machine. Don’t scan untrusted third-party code without isolation.

Can I use it for private repositories?

Yes. Set a GITHUB_TOKEN environment variable with repo access, and Open Kritt can clone and scan private GitHub repositories. Your code stays on your infrastructure—nothing is sent to Kritt’s servers.

Does it integrate with CI/CD?

Open Kritt provides a CLI and API, so you can integrate it into CI pipelines. Run scans on PRs, block merges with high-severity findings, or generate security reports on each release. The workflow system makes it easy to define consistent checks.

What models work best?

Claude models (especially Claude Sonnet and Opus) and Codex tend to perform well on security research tasks due to their code understanding and reasoning capabilities. You can experiment with different providers via OpenRouter to find what works best for your codebase.

How do I create custom workflows?

Workflows are sequences of prompt steps defined in the UI. Each step specifies what to analyze, what to look for, and how to handle the output. You can chain steps (output of one feeds into the next), run steps in parallel, and add post-scripts for validation. See the workflows documentation for details.