One line of buzzwords turned into 200 stars: how a small Barcelona team's AI pentest agent got amplified

Vulnerator-AI is an open-source project for "AI doing pentesting on its own". It's three weeks old and has 268 stars. The numbers aren't big. The top projects in this space have tens of thousands of stars, but the way this one grew is clean, and every step is visible: priming in the native-language community first, a one-line repo description with buzzwords built to be reposted, then waiting for a big security-circle account to share it verbatim. If you want your open-source project seen, this playbook can be copied. If you want to build yet another AI pentest agent, it isn't worth it.

A tweet that copied the bio word for word

On Oct 29, 2026 13:06 (UTC), Nicolas Krassas (@Dinosn on X, an old account in security circles that shares tool links every day) posted a tweet whose body was this repo's bio, not a single word changed:

Autonomous OODA-loop pentesting agent. Hermes brain, Kali + Playwright tools over MCP, ATT&CK as the Orient language.

followed by a GitHub link. The tweet got 527 likes, 83 reposts, and about 21k views (X, 11-06).

Before that, the project had been up for two weeks with 40 stars in total. GitHub star counts over the next two days: 10-29, then 40,10-30, then 73,10-31, then 164, a gain of 91 in one day. By 11-06 it was at 255 to 268 stars (two counting methods), with 62 to 64 forks. It never reached GitHub Trending, and I couldn't find any Hacker News or Reddit threads discussing it. Nearly all of the growth lines up with that one tweet.

What it is

One Docker Compose file, five containers: a Kali box loaded with nmap, nuclei, sqlmap, ffuf and Prowler; a headless browser (Playwright); Caido, a web security tool; a "brain"; and an MCP entry point so tools like Claude Code and Kiro can call it remotely. The brain is Nous Research's Hermes Agent, which by default calls Mistral Large through OpenRouter. It runs on an OODA loop (Observe, Orient, Decide, Act, a decision model that started with fighter pilots), on its own: scanning ports, reading web pages, picking vulnerabilities, trying them out, and labeling each step with MITRE ATT&CK categories.

The README is blunt about it: "Deliberately minimal: no core, no REST API, no UI, no gate." Setup takes one OpenRouter key and ./install.sh. It also ships a prompt: "Don't want to read the docs? Paste this into Claude Code or Cursor and let it install for you." Practice targets like OWASP Juice Shop can be spun up alongside.

It was built by AKTIOS, a security firm in Barcelona, and its author, Sergio Jiménez Narbona (GitHub account irraz). Three weeks, 9 releases, from v1.0.1 (10-16) to v1.8.0 (11-05).

The two weeks before the repost: Spanish LinkedIn

Before it got reposted, the project circulated in Spanish-language LinkedIn circles. On the day the repo was created (10-15), the author posted the first message: "Continuous AI-orchestrated pentesting is here, and it's here to stay." A few days later his colleague Juan Holgado posted one that traveled further:

"We just built an AI with military knowledge and gave it a Kali box 😈 Meet Vulnerator-AI. No, it's not a Norwegian death metal band (though the team T-shirts are on the way 🤘)... It's like hiring a retired soldier to do continuous pentesting on Kali: observe, orient, decide, act, on repeat."

That post got 52 reactions and 16 comments, mostly from security practitioners in Spain and Latin America. The author then posted a few more release announcements, and the company account reshared them.

During the same stretch, the author submitted PRs and issues to several awesome lists (awesome-ai-pentesting, Awesome-AI-Hacking-Agents, awesome-security-agent-harnesses), and the project was listed in directories like Infosec House and DeepWiki. In October 20, No cON Name, Barcelona's long-running hacker conference, published its schedule: on November 26–27, the author was leading a hands-on "Hands-On Pentesting-AI" session teaching this project, and the conference itself forked the repo and built a site for it.

Together, these brought in 40 stars over two weeks, roughly the ceiling a native-language circle can deliver. But by the time a big account noticed it, the project already looked like something that ships often, has a logo, a demo video, and conference backing.

Why the big account reposted it as-is

Inference: accounts like @Dinosn repost many links a day, and the ones they repost most are things that need no explanation, that readers grasp at a glance, and that feel new. Vulnerator's one-line description packs in terms that people in five overlapping circles (security, AI, and more) all know: OODA, Hermes, Kali, MCP, ATT&CK. Hermes Agent was one of the hottest open-source agent frameworks of 2026 (261k stars on its own), and the badge at the top of the README reads "Hermes Agent — Nous Research". I couldn't find how he came across the project.

After the repost, the afterglow was all secondhand: a Mastodon user wrote a long commentary (10-29); an auto-generated tech site published a write-up (10-30); GitHub Signals, an AI-run short-video account, posted a clip called "a pentester that thinks for itself" on TikTok, Instagram and YouTube Shorts (starting 10-30); and 11-01, Japanese security writer Takashi Yoshikawa (@MalwareBibleJP, author of Malware Textbook), published a detailed Japanese explainer with 44 likes and 2,760 views. These pushed stars slowly from 164 to 255. The author's own posts on X barely travelled: 11-06 his v1.8.0,13 post got only a few views.

What actually drove the stars was borrowed buzzwords and that one repost; technically it isn't stronger than its peers. One-click install, practice targets and polished visuals handle a different job: they decide whether people who land on the page star it and actually pull it down and run it. The fork-to-star ratio reached 24%, which is on the high side for open-source projects.

It's in a crowded lane

Comparable open-source projects (GitHub, 11-06):

On the commercial side there are paid enterprise products like XBOW and Horizon3 NodeZero. "AI autonomous pentesting" already has players on both the open-source and commercial sides with tens of thousands of stars and funding. Vulnerator's 268 stars came from the edge of that red ocean, riding Hermes's hype. I haven't seen anyone copy it; the category was already crowded, so there's no question of who copied whom.

If it were you

The playbook is copyable; the category isn't. "Trending upstream framework + one vertical + a one-line buzzword description + native-language pre-launch + a big account reposts it" costs one or two people three weeks and can earn a few hundred stars and a name in the scene. But another AI pentest agent will just sit behind a pile of projects with tens of thousands of stars.

Who it suits: developers or small teams with real expertise in a vertical (security, ops, data, law, design all count), a native-language professional community, and the ability to ship consistently for three weeks. What they want is credibility, speaking slots and standing in the scene, not tens of thousands of stars.

Who it doesn't suit: anyone building yet another general-purpose AI pentest or AI coding agent (the top projects have tens of thousands of stars, and no README will squeeze you in); anyone with no community who plans to post once and hope; anyone expecting star counts to turn into money.

How to do it: pick an upstream agent framework that's gaining momentum (today Hermes Agent, but it could be the next one) and apply it to a vertical you know well, ideally one that doesn't yet have a few thousand stars of competitors. Design by subtraction: one key, one install command, one prompt to "paste into Claude Code and let it install it for you", one practice environment that runs immediately. Write the repo description as a single line with three to five terms that both your community and the AI crowd recognize, written so people can repost it verbatim. Have a logo, a demo video and badges, so a three-week project looks like a six-month one.

How to get users: start in your native professional circle (LinkedIn, Jike, Qiita…), with colleagues and the company account resharing, and a memorable hook ("Give AI a Kali box"). At the same time, submit PRs to relevant awesome lists and get listed in tool directories, since big accounts and AI aggregators often find things there. Then look for accounts in your vertical that specialize in reposting tools on X. That step is partly luck, but the more entry points you have, the more likely you are to be seen. A workshop tied to an offline conference is even better, since it gives you a reason to keep shipping releases.

Timing: buzzwords have a shelf life. Build while Hermes's hype is still alive; my estimate is it lasts into the first quarter of 2027. Once repos called "Hermes + [some domain]" multiply, this one-line description stops feeling new.

Signs you've misread this: (1) Vulnerator's stars keep climbing clearly after the late-November No cON Name conference (for example, past 1,000 by late December), which would mean there's still room for small projects in the AI pentest lane and it's worth reassessing; (2) conversely, if a wave of "Hermes + vertical" projects appears in the coming months and most stall at a few dozen stars, then the real driver was that one @Dinosn repost, not a playbook you can replicate.