open·kritt organizes AI-assisted security reviews

What this is about

open·kritt is an open, self-hostable platform for security researchers and developers who want AI agents to inspect source code for vulnerabilities. The project was published on GitHub in July 2026 under the AGPL-3.0 license. Instead of handing an entire repository to one model with a broad request, the tool splits the analysis into smaller, clearly scoped steps.

That approach is useful because automated security reviews often have two opposite problems: they generate many unusable alerts or miss flaws that require several steps to prove. open·kritt aims to connect analysis, validation, prioritization, and reporting in a workflow that people can inspect.

What open·kritt actually does

Users build reusable workflows from focused prompts. Codex or Claude Code can execute these steps in parallel. Targets can be local or remote repositories and their dependencies. The platform then combines results, removes duplicate findings, and applies a configurable severity ranking.

Validation scripts can run tests, compile targets, and produce evidence for a vulnerability. Model access can come from a Codex login or connections to OpenAI, Anthropic, and OpenRouter. The interface runs locally, and the documented setup requires Git, Docker Compose, and Node.js 20 or newer.

One technical boundary matters: according to the project documentation, the backend has no built-in application authentication. Default ports bind to 127.0.0.1, but the system must not be exposed to the public internet without additional protection.

Why it matters

A useful security finding is more than a suspicion. Teams need to locate affected code, test exploitability, merge related alerts, and hand developers a comprehensible report. General chat assistants often lose context at these transitions.

open·kritt makes the steps visible and repeatable. This is relevant to AppSec teams, bug-bounty researchers, and developers who want control over prompts, models, and infrastructure. The AGPL license permits inspection and self-hosting, but network-based modified deployments may require the changed source code to be offered to users. Organizations should review that obligation before adoption.

The vendor also points to more than $1.5 million in bug-bounty payouts earned by the research team under the Blockian name. That self-reported background indicates practical experience, but it is not an independent measurement of the product's effectiveness.

In plain language

Think of open·kritt as a workshop inspection for a car. One mechanic does not merely listen to the engine for a moment and write “possibly broken.” Several defined stations inspect brakes, electronics, and fluids, confirm anomalies, and merge duplicate reports into one finding.

The AI agents are the inspection stations. The human security researcher remains the foreman who defines the process, limits dangerous tests, and decides whether a finding is credible.

A practical example

A team maintains a web application with 250,000 lines of code and wants to review authentication before a release. It creates a five-step workflow: map the attack surface, inspect session logic, rank suspicious locations, run reproducible tests, and document confirmed findings.

Four agents handle separate areas. They initially return 32 alerts. Deduplication reduces related reports to twelve candidates. Validation scripts confirm three flaws; nine remain unproven or turn out to be false positives. Only then does an experienced person review the evidence and decide priority and remediation. These numbers are fictional and are not a published performance metric for open·kritt.

Scope and limits

First, tool-enabled agents run as root in disposable job containers. They receive writable repository copies and direct internet access. A malicious repository or faulty command can therefore cause significant harm. The developers recommend a dedicated Docker host or virtual machine.

Second, an AI finding is not proof of an exploitable vulnerability. Models can misunderstand data flows, run incomplete tests, or produce convincing false positives. Disclosure, ticket priority, and production changes require human review.

Third, operation takes work and consumes model resources. Teams must maintain prompts, protect credentials, monitor containers, and compare results. For an occasional scan of a small codebase, established static-analysis tools may be faster. open·kritt also replaces neither conventional SAST and dependency scanners nor manual penetration tests.

The sensible next test is therefore small: use a non-production repository copy inside an isolated VM, run one tightly scoped workflow, and compare confirmed findings with existing scanners.

SEO & GEO keywords

open·kritt, Kritt AI, AI security review, AppSec, vulnerability research, security agents, self-hosted AI, Codex, Claude Code, AGPL-3.0, Docker, code security