OpenHunterAI is a local-first workspace that combines browser inspection, recon, ZAP, and a Nuclei adapter with attacker-style AI reasoning to assess web, API, and LLM applications you own or are authorized to test.
| Tool | LumosLab-Innovation/OpenHunterAI — local-first AI red team workspace for web, API, and LLM application security |
| Category | AI-assisted application security assessment platform |
| Primary Use | Scope-verified, human-gated security assessment of web, API, and LLM applications from an external attacker's perspective, without source code |
| Safe Use | Explicitly designed for authorized testing: domain verification, scope locks, and approval gates constrain every run to targets you own or are contracted to assess; unsuitable for third-party scanning |
| Telemetry Note | All activity — scan plans, worker activity, sanitized findings — is stored in a local Postgres database; the workspace avoids retaining raw sensitive traffic, and its bounded adapters avoid destructive actions, making runs observable primarily in your own logs and the target's own access logs |
OpenHunterAI positions itself as "your local AI red team," and the framing is more disciplined than the tagline suggests. Rather than an autonomous attacker, the project is a self-hosted workspace that brings scope management, scan activity, findings, and remediation guidance into a single local database, with no signup and no cloud dependency beyond the model providers you configure yourself. Written primarily in TypeScript with Go workers behind it, the repo sits at a respectable 316 stars and is tagged across appsec, owasp, pentesting, and ai-security topics. The assessment philosophy is black-box: it assesses externally observable behavior from an attacker's perspective and explicitly does not require source code, which places it squarely in the external authorization-and-verification workflow of a professional engagement rather than a code-audit tool.
What distinguishes the project architecturally is its insistence that authorization comes before capability. The assessment model is a six-stage pipeline: lock scope, map surface, establish baseline, test bounded hypotheses, gate evidence, then report and retest. Each stage has an explicit control attached — scope lock requires authorization before work begins, hypothesis testing refuses destructive actions and gates sensitive validation behind human approval, and the evidence stage separates observed signals from analyst judgement. This is a governance-first design that many AI security tools skip entirely, and it is the strongest signal in the README that the authors understand the operational liability of an AI-driven scanner pointed at someone else's infrastructure.
Scope enforcement is concrete rather than aspirational. Before any scanning happens, you create a project and verify its public domain; the tool then constrains every run to that verified target and the approved paths within an agreed scan plan. The README is also blunt about a common misreading: "local" means the workspace runs on your machine, not permission to scan localhost or private networks, and it instructs you to keep its ports on loopback. Test-surface definition can be enriched with OpenAPI/Swagger descriptions, Postman examples, sanitized HTTP captures, and dedicated test accounts — while the project explicitly disclaims automated import or retention of raw sensitive traffic as a capability, a privacy posture worth noting.
The signal-gathering layer is a hybrid of classic tooling and AI reasoning. Included out of the box are browser inspection, recon, OWASP ZAP integration, and a Nuclei adapter, with optional interactive target-login capture via a browser flag. The AI-assisted investigation side is deliberately pluggable: the README describes OpenHack and Strix adapter boundaries where you bring your own runtimes rather than the tool shipping an opaque model integration. It is honest about current gaps — the Nuclei adapter is packaged but a reviewed template bundle is not yet supplied, so users must resolve a template-policy blocker before scanning. That kind of candid "known limitations" section is rare and suggests a project managed with release gates rather than hype.
Under the hood, the architecture is a small distributed system orchestrated through Docker Compose. A React + Vite frontend talks to a public API that builds the scan plan; Go workers execute that plan through bounded adapters, communicating over a NATS message bus; an internal API handles findings and reporting, persisting everything to Postgres. Reporting is a backend service, not a scan worker, which means findings pass through a findings policy before entering reports — a deliberate sanitization boundary between raw signals and what a human eventually reads. The separation of plan-building from execution is what makes the approval-gate model technically enforceable rather than merely a UI checkbox.
Installation is refreshingly contained. With Git, Node.js 22+, and Docker Compose v2 in place, git clone https://github.com/LumosLab-Innovation/OpenHunterAI.git followed by node ops/local.mjs start builds the images, applies database migrations, and generates a private .env.local with fresh secrets; the workspace then listens on localhost:3001. You set provider keys and model IDs in .env.local, restart, verify a domain, authorize the scope, and only then scan. Operational commands like node ops/local.mjs status, logs public-api, and stop (which preserves database volumes) round out a lifecycle that stays entirely on your machine.
A second, more novel integration path targets coding agents directly. A portable skill can be installed into Codex, Claude Code, Gemini CLI, Grok Build, and OMP via npx --yes skills add LumosLab-Innovation/OpenHunterAI --skill openhunter, giving your agent a read-only API helper and a workflow that prepares a scoped assessment but does not start a scan on installation. The suggested prompt pattern — check my workspace, prepare a scoped assessment, wait for approval before scanning — mirrors the tool's human-in-the-loop philosophy. For teams already living inside agent-driven development loops, this is a pragmatic way to fold pre-engagement reconnaissance and scoping into the same environment where the code is written, even though the assessment itself remains black-box.
Licensing deserves attention before adoption. The repo is source-available under PolyForm Noncommercial 1.0.0 (the metadata reports NOASSERTION, consistent with a nonstandard license file), with commercial licensing handled separately in LICENSING.md. That makes OpenHunterAI free for research, learning, and internal use but a legal review item for any consulting firm wanting to run it on client engagements. Upstream tools and their notices are documented in docs/UPSTREAM_SOURCES.md, so attribution hygiene appears to have been considered. The documentation tree is unusually deep for an alpha — architecture, worker specifications, LLM provider routing, security guardrails, and acceptance criteria all have dedicated documents.
The project labels itself alpha, and the README's own caveat is the right lens for evaluation: "a healthy workspace or an empty report does not prove a target is secure." Integration coverage depends on which runtimes you configure, and the absence of a reviewed Nuclei template bundle means the flagship template scanner needs operator curation before it produces meaningful results. From a defensive perspective, the tool's bounded adapters, no-destructive-action policy, and sanitized evidence pipeline make its activity comparatively legible in target-side logs — standard scanner traffic patterns from ZAP and Nuclei fingerprints, originating from the operator's own egress, against a scope the target owner has presumably agreed to.
For authorized professionals, OpenHunterAI is best understood as an orchestration and governance layer over familiar offensive tooling, with AI reasoning applied at the hypothesis and triage stages rather than as a free-roaming agent. Its value proposition is the workflow: verified scope, deterministic plans, human gates, sanitized findings with coverage gaps, fix guidance, and manual retest built in. Teams evaluating it should weigh the noncommercial license, the alpha maturity, and the need to supply their own reviewed Nuclei templates and optional OpenHack/Strix runtimes — but the bones of a genuinely accountable AI assessment platform are visible in the design.
LumosLab-Innovation/OpenHunterAI.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.