OpenHunterAI releases an AI red-team engine after killing the startup

Why it matters

OpenHunterAI shows one path for abandoned AI startups: release the core product for distribution, retain commercial rights and let adoption test whether a business remains.

Nicolas Krassas (@Dinosn) said in a post on X on September 14th that his group had abandoned plans for an AI security startup and released its red-team engine as OpenHunterAI.

"We killed the startup idea and released the red-team engine instead," Krassas wrote. The decision puts the code in developers' hands while preserving the option to charge for commercial use, a narrower bet than building and selling another hosted security platform from scratch.

Krassas brings an operator's background to that decision. His public profile lists threat and vulnerability management work at Henkel, penetration testing and security leadership roles, along with earlier work securing the official Athens 2004 Olympics website. He also studied neural networks at the University of Leeds, giving the project a technical lineage that predates the current rush to attach language models to application-security scanners.

OpenHunterAI runs on a user's machine through Docker Compose and is designed to assess verified public websites, APIs and LLM applications. Users define a target, verify its domain, authorize a scope and approve a scan plan before workers begin testing. The interface collects scan activity, evidence, findings, remediation guidance and manual retests in one local workspace.

The product combines browser inspection, reconnaissance, OWASP ZAP and an adapter for the Nuclei vulnerability scanner. AI models assist with triage, attack hypotheses and remediation guidance. Users supply their own model-provider credentials, and the documentation cautions that local hosting does not make the system fully offline: requests can still reach external model providers and the target under assessment.

OpenHunterAI also includes a skill that lets coding agents prepare assessments from inside an application repository. The published installation command names Codex, Claude Code, Gemini CLI and Grok Build as supported clients. The skill starts with a read-only helper and requires human approval before a scan begins.

The release is an alpha, with blockers documented in the repo

The public repository is unusually direct about what remains unfinished. Its release-readiness review labels OpenHunterAI an alpha and says the review covers source code rather than evidence from a fresh production deployment or a complete target scan.

One blocker concerns Nuclei. OpenHunterAI packages an adapter, but the reviewed template bundle and its configuration are incomplete. The review warns that process errors can appear as an empty successful result because one execution path parses output without fully handling the underlying process failure. OpenHunterAI's own documentation tells users to resolve that blocker before scanning.

OpenHack and Strix, two other security runtimes referenced by the project, must also be installed separately. The readiness review calls for a complete authorized end-to-end scan, including callbacks, reports, exports and manual retesting, after the blockers are fixed. It also says the PDF export truncates long reports, making JSON, DOCX or HTML the safer formats for complete output.

Those caveats matter for a tool that can otherwise produce the appearance of security coverage. OpenHunterAI explicitly warns that a healthy workspace or an empty report does not prove a target is secure. Its design separates observed signals from analyst judgment and leaves impact decisions and closure with a human reviewer.

The project's security rules prohibit private-network scanning, destructive actions, denial-of-service testing, persistence, credential stuffing and raw-secret exfiltration. Potentially mutating requests, access-control checks, payment actions and high-severity retests require explicit approval. Sensitive traffic and credentials are also barred from model prompts, reports and stored evidence.

Source-available, with commercial rights retained

Despite the "Open" in its name, OpenHunterAI is source-available rather than open source under the standard industry definition. Its licensing file uses the PolyForm Noncommercial 1.0.0 license and explicitly says it is not approved by the Open Source Initiative.

Personal research, education, experimentation and other noncommercial uses are permitted. Paid services, commercial hosting, resale and inclusion in a paid product require a separate license from Henry Don, identified in the repository as the copyright holder and commercial licensing contact.

That structure reveals the strategy beneath the abandoned startup. Releasing the engine can attract users, testing and integrations without requiring OpenHunterAI's creators to finance a full sales operation, hosted service and enterprise support organization. The retained commercial rights still leave room to license the software if companies decide to put it into production or sell services around it.

For now, OpenHunterAI is best read as working product code under public inspection, paired with a candid list of unfinished release work. The startup pitch is gone. The underlying commercial option is still there.