On March 2, 2026, a developer named Yohannes Gebrekirstos released an open-source framework called BlacksmithAI: a system that lets AI agents run a full penetration test on their own, the kind of structured security assessment that normally takes a trained human, or a small team of them, working through several distinct phases in sequence. Help Net Security covered the launch with direct comment from the creator; a handful of security-trade outlets and dev-community blogs picked it up afterward. All of that coverage is aimed at security professionals. This explains the same thing for anyone who isn’t one.

What It Actually Is

BlacksmithAI is a hierarchical multi-agent system, according to its own documentation. One orchestrator agent manages the overall plan and delegates work to five specialised subagents: a reconnaissance agent that maps a target’s attack surface, a scanning and enumeration agent that probes discovered services in detail, a vulnerability analysis agent that maps findings to known weaknesses and prioritises them by risk, an exploitation agent that attempts controlled proof-of-concept exploits, and a post-exploitation agent that assesses what damage a successful breach would actually cause. The orchestrator compiles all of it into a final report with findings, severity ratings, evidence, and remediation guidance.

Each agent works inside a shared container called “mini-Kali” — a pre-configured environment loaded with the tools that agent needs, rather than spinning up a fresh container for every task. According to the creator, this was a deliberate efficiency choice: agents share one environment with tool caching and access controls that stop them from adding or removing tools mid-run, keeping results consistent between attempts.

The agents’ reasoning runs on a large language model, which can be a cloud provider like OpenRouter or a model hosted locally through vLLM, so the project doesn’t lock anyone into one AI vendor. It’s available through a command-line interface or a web UI, runs on Linux, macOS, or Windows via WSL2, and needs Docker plus at least 4GB of RAM to operate. The code is released under the GPL-3.0 license, with a separate commercial license available for anyone wanting to use it in closed-source products.

Why It’s Built as a Team of Agents, Not One

The design choice that actually distinguishes BlacksmithAI from other AI security tools isn’t any single feature — it’s the division of labour. Most existing AI security tools rely on a single agent trying to handle every phase of an assessment. The creator’s stated reasoning, given directly to Help Net Security, is that real penetration testing is normally a team effort, with specialists handling different phases rather than one person — or one AI — doing everything at once. The orchestrator acts as the process lead, breaking a complex task into sub-goals and handing each one to whichever specialist agent is built for it, the same way a human pentest lead would divide work across a team.

It’s also explicitly unfinished in some ways worth knowing about. The creator has said future versions aim to add browser interaction — agents currently can’t click buttons or fill out forms, only discover paths and known issues through tools like Nikto and Gobuster — along with support for interactive frameworks like Metasploit and BeEF, integration with external services like Shodan through the Model Context Protocol, and a “skills” system letting agents learn best practices by combining multiple tools rather than using them in isolation.

What’s Actually Inside the Toolkit

The mini-Kali container bundles 28 named, established, professional-grade security tools — the same category of software a human penetration tester would reach for, just orchestrated by AI agents instead of typed by hand.

Who Actually Benefits From This

The honest answer is a narrower group than the framing of “AI does your security testing now” might suggest. A few categories of people get genuine value out of a tool like this existing, and a few don’t.

What It’s Not

It’s not a one-click attack button, and it’s not the only project doing this — it’s one open-source entry in a wider, fast-moving category of AI-driven offensive security tooling that emerged through 2026. As of writing, the project has 222 GitHub stars and 39 forks — a real, growing community, but still a young one.

My Take — Mr Wangdoo

Reading through the documentation and the creator’s own comments, what stands out isn’t really the AI angle — it’s that this is fundamentally a labour-division problem someone solved well. Multiple narrow agents instead of one generalist doing everything is a sensible engineering choice, and it’s the same lesson plenty of software has learned long before AI agents existed; BlacksmithAI just applies it to a domain, offensive security, where the tools being orchestrated are unusually high-stakes.

What I find more interesting than the tool itself is what its existence signals: the orchestration layer that used to require a trained human coordinating several specialised tools by hand is now something a single open-source project can package and give away under GPL. None of the individual tools inside it are new. The packaging is.