Listen to this Post
Introduction:
The traditional approach to AI-powered vulnerability scanning—feeding an entire codebase to a large language model and asking it to “find everything wrong”—has consistently produced overwhelming false positives and missed critical business-logic flaws. Open-Kritt, an open-source, self-hosted security research platform from the team behind $1.5 million in bug-bounty earnings, fundamentally reimagines this paradigm by orchestrating specialized AI agents that operate like a coordinated security team, each tackling discrete, well-defined tasks in isolated containers.
Learning Objectives:
- Understand the multi-agent orchestration architecture that enables parallelized, focused vulnerability discovery across codebases of any size.
- Learn to deploy and configure Open-Kritt with major AI providers including OpenAI, Anthropic, and OpenRouter.
- Master the workflow of entry-point mapping, path tracing, dynamic analysis, and automated proof-of-concept generation.
- Implement containerized agent isolation and understand the security implications of running AI agents with system-level access.
You Should Know:
1. The Multi-Agent Architecture: Breaking the Monolith
Instead of issuing a single, monolithic prompt to an AI model, Open-Kritt decomposes security research into a pipeline of focused tasks executed by parallel agents. The workflow begins with a reconnaissance agent that maps every reachable entry point in the target repository—identifying API endpoints, user inputs, public methods, and external interfaces. Once the attack surface is mapped, the system launches separate agents to trace each code path independently, searching for specific classes of vulnerabilities including injection flaws, authentication bypasses, privilege escalation vectors, and cryptographic misuses.
Each agent operates within its own isolated container, provisioned with the ability to inspect source code, install security tools, compile the project, run test suites, fuzz inputs, and even build functional proofs of concept. This containerized approach ensures that agents cannot interfere with one another or compromise the host system, even when analyzing untrusted or malicious code. The raw findings from all agents are then aggregated, de-duplicated when they describe the same bug, and ranked by severity so security researchers can prioritize the most critical issues first.
Step-by-Step: Deploying Open-Kritt
Prerequisites: Git, Docker with Docker Compose, and Node.js 20+ git clone https://github.com/Kritt-ai/open-kritt cd open-kritt Run the setup wizard - guides through API key configuration ./kritt setup Start the full stack (backend, worker, and frontend services) ./kritt start Access the web interface Open http://localhost:5173 in your browser
The setup wizard supports authentication via Codex login, OpenAI API keys, Anthropic API keys, or OpenRouter. A `GITHUB_TOKEN` is optional and only required for scanning private repositories. By default, all services bind to 127.0.0.1, and the backend does not include application authentication—the platform should be kept private and run on a dedicated Docker host or VM.
2. Agent Isolation and the Threat Model
Tool-enabled agents run as root inside disposable job containers with writable repository copies and direct internet access. This design enables agents to install dependencies, compile targets, run dynamic analysis tools, and build proofs of concept without restrictions. However, this power comes with significant security considerations: an agent compromised by malicious code could potentially escape its container or abuse network access.
Open-Kritt addresses these risks through several mechanisms. Each job runs in a fresh container that is destroyed after task completion, preventing persistence across analyses. The platform includes a documented threat model and recommends deployment on isolated infrastructure. For production deployments, additional hardening measures should be implemented:
Example: Restrict container capabilities and enforce seccomp profiles docker run --rm \ --cap-drop=ALL \ --cap-add=NET_RAW \ --security-opt=seccomp=/path/to/seccomp-profile.json \ --read-only \ --tmpfs=/tmp \ kritt-agent:latest
3. Integrating with AI Providers and Security Tools
Open-Kritt supports multiple AI providers through a unified interface, allowing teams to leverage the strengths of different models for different tasks. Codex and Claude Code excel at reasoning about business logic and data flows, while OpenAI’s models and Anthropic’s Claude provide strong general-purpose code analysis capabilities. OpenRouter enables access to a broader ecosystem of models through a single API.
The platform’s extensible architecture allows security researchers to define custom workflows—chaining focused prompts into reusable security research playbooks. Post-scripts can validate findings, build proofs of concept, and produce detailed reports. Custom severity rankers can be applied to prioritize results according to organizational risk thresholds.
4. Real-World Validation: $1.5M in Bug Bounties
The Kritt team, operating under the researcher name Blockian, has earned over $1,500,000 in bug-bounty payouts across platforms including Immunefi and HackenProof. Open-Kritt is the open-source distillation of the internal system that generated these earnings. In nine months of internal use, the system discovered more than 20 High and Critical vulnerabilities. Most notably, it finished first in the Firedancer V1 audit competition—a $1,000,000 bug bounty challenge targeting Solana’s 636,000-line C codebase—using fully AI-driven research with no manual review.
5. Validation and Verification Workflow
Raw AI-generated findings require validation to eliminate false positives. Open-Kritt addresses this through a multi-stage verification pipeline:
- Post-script validation: Custom scripts automatically test each finding against the live codebase.
- Proof-of-concept generation: Agents attempt to build working exploits that demonstrate the vulnerability.
- Deduplication: Findings describing the same underlying issue are merged into a single report.
- Severity ranking: Custom rankers apply organizational risk scoring to prioritize remediation.
For Windows environments, validation scripts can be adapted using PowerShell:
Example: Validate a potential path traversal vulnerability
$targetPath = "....\windows\system32\drivers\etc\hosts"
try {
$content = Get-Content -Path $targetPath -ErrorAction Stop
Write-Host "[!] Path traversal confirmed - file accessible"
} catch {
Write-Host "[+] Path traversal not exploitable"
}
6. Security Hardening for Production Deployments
When deploying Open-Kritt in enterprise environments, additional security controls should be implemented:
docker-compose.override.yml - Production hardening services: worker: security_opt: - seccomp:./seccomp.json - apparmor:docker-kritt cap_drop: - ALL cap_add: - NET_RAW read_only: true tmpfs: - /tmp:rw,noexec,nosuid,size=2g networks: - isolated environment: - KRITT_MAX_CONCURRENT_JOBS=2 - KRITT_JOB_TIMEOUT=3600
Network isolation is critical—workers should not have access to internal production networks or sensitive resources. Rate limiting and job quotas prevent resource exhaustion attacks. Regular updates to the base container images and AI provider SDKs mitigate known vulnerabilities.
What Undercode Say:
- Key Takeaway 1: Open-Kritt represents a paradigm shift from monolithic AI prompts to orchestrated multi-agent security research. By decomposing vulnerability discovery into focused, parallel tasks, it achieves accuracy and coverage that single-prompt approaches cannot match. The platform’s containerized agent architecture provides both isolation and flexibility, enabling agents to install tools, compile code, and build proofs of concept without compromising the host system.
-
Key Takeaway 2: The platform’s provenance is its strongest validation—the Kritt team has earned $1.5 million in bug bounties using the internal version of this system. Winning the Firedancer V1 audit competition with fully AI-driven research demonstrates that agentic approaches can outperform manual review in complex, large-scale codebases. For security teams and independent researchers, Open-Kritt offers enterprise-grade AI-powered vulnerability discovery without the cost and lock-in of commercial solutions.
Prediction:
-
+1 Open-Kritt will accelerate the commoditization of AI-powered security research, enabling smaller teams and independent researchers to compete with well-funded security firms. The open-source model will foster a community of contributors extending the platform with new agent types and detection capabilities.
-
+1 As AI agents become more capable of autonomous code execution and dynamic analysis, the line between automated vulnerability discovery and automated exploitation will blur. Expect increased focus on guardrails, ethical use policies, and responsible disclosure frameworks around agentic security tools.
-
-1 The democratization of advanced AI-powered vulnerability discovery will lower the barrier to entry for malicious actors. Threat actors will adopt similar multi-agent architectures to accelerate zero-day discovery, forcing defenders to adopt AI-powered detection and response at an accelerated pace.
-
+1 The success of Open-Kritt in the Firedancer audit signals a broader trend toward AI-driven code audits becoming the industry standard. Organizations will increasingly require AI-assisted security reviews as part of their SDLC, driving demand for platforms that can orchestrate multiple AI models and tools in a unified workflow.
🎯Let’s Practice For Free:
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
IT/Security Reporter URL:
Reported By: Iamsangamofficial A – Hackers Feeds
Extra Hub: Undercode MoN
Basic Verification: Pass ✅
