Earlier this week, a repo called "google-ultra" was #1 in Python on my own site. It promised paid Google AI tiers for free. I blocked it in the morning, and GitHub took it down within a day.

It had climbed my ranking the way the rest of these campaigns do: with stars delivered in batches, to brand-new accounts, on the same days. This post is what I've seen this month, with the numbers, the mistakes my filters made, and how you can spot these repos yourself.

Why a growth ranking is a target

I build Upstar alone, from Abidjan. It ranks the GitHub repos that are taking off in 69 topics, every hour. The score is the stars a repo gained in the period, divided by the square root of the stars it had before. A small project that doubles in a week can beat a giant that gains 2,000.

That's the whole idea, and it's also the weak spot. Fast growth is exactly what a fake project tries to manufacture. A few hundred cheap stars on a brand-new repo can push it into the top 10 of a niche topic.

So every repo goes through filters before Upstar shows it anywhere: rules, a small classifier I trained on past decisions, and LLM judges that read the README and the files. Since October, they have blocked more than 3,800 repos.

8 stars, then 17 stars

Upstar keeps a day-by-day star count for every repo it ranks. When I lined them up this week, one pattern jumped out.

40 repos, owned by 40 different accounts with zero followers, all received exactly 8 stars on November 14. Three days later, on November 17, all 40 received exactly 17 stars. 25 of them had also received 3 to 5 stars on November 12. 38 of the 40 accounts were created from October 27 onward, most of them in batches.

Different owners, different repos, identical deliveries on the same days. That's not 40 projects getting lucky. That's one seller filling one order.

One caveat, because I almost got it wrong. Seven real projects also gained exactly 8 and 17 stars on those days, by pure chance: audiobookshelf, newsnow, mlx-audio and others, between 1,000 and 22,500 stars. Their curves are smooth, with stars coming in every day. The numbers alone prove nothing. The numbers on dozens of brand-new, empty accounts do.

Real project names, no real code

Some of these repos borrow the name and description of a real open-source project and add a prefix. One was called resp-expandModeIntegerScaler, a copy of a real Godot addon. Another copied a retro game project the same way.

The code isn't there. Each of these two repos holds just two files: an HTML page and a CONTRIBUTING.md. The README is empty, so the bait sits in the contributing file instead:

  • a big button to a download on an outside site;
  • a password for the archive;
  • and the line that should make anyone stop: "Run as Administrator".

The prefix "resp" shows up in the repo names of 136 young accounts. Another 66 accounts were created on October 28 within 35 minutes, many of them with repo names built on the same few words.

A cheat network built in 47 seconds

The second campaign sells cheats for an online game called Rob Run. On October 14, six GitHub accounts were created between 13:23 and 13:40 UTC, five of them within 47 seconds. Each account created its repo about two minutes later.

Every README tells you to run robrun-hk.exe. That file isn't in the repo. It comes from a zip uploaded by hand to the releases, and the code in the repo doesn't build it.

They were careful in small ways:

  • each zip has a different size, from 1.65 to 2.92 MB, so they don't look like copies of one file;
  • the repos hold 9 to 16 real-looking code files, so they don't look empty;
  • the forks were bought too: 41 of one repo's 80 forks came from accounts all created on November 7.

One copy of the same cheat went further. Its button ran a PowerShell one-liner that fetched a script from a GitHub Pages site, and the script decoded a payload with base64 and XOR. I haven't analyzed what it installs.

What my filters missed

I'd rather tell you where Upstar failed than pretend it didn't.

The executable rule only looked at near-empty repos. I flagged "run this .exe" only when a repo had almost no code. The Rob Run repos had enough files to slip under it. Now a game cheat that asks you to run a program the repo doesn't contain is blocked, however much code sits next to it.

The shared-file rule wanted identical files. Six zips of six different sizes passed as six different files.

The judges were too polite about cheats. My policy kept online-game cheats out of the spotlight without blocking them. The judges saw real code and a zip hosted on GitHub, and they kept them. That's fixed in the policy now.

Old approvals never got a second look. Two of the "resp" repos were approved on November 6. At the time, the judges didn't read the CONTRIBUTING.md when the README was empty, so they never saw the bait. Approvals made before that change are now re-read.

The classifier knew, but nobody asked it. My classifier scored the Rob Run repos up to 0.998 out of 1. But it could only confirm a block that the first judge proposed. When the judge said "keep", the score went nowhere. Now any repo it scores at 0.9 or more goes to a stronger model for a decision.

What did work: since November 18, that classifier has blocked 268 lures on its own, and none of those blocks has been reversed so far.

How to spot one yourself

None of these signs is proof on its own. Two or three together should make you close the tab.

  1. The account is as young as the repo. Created the same day, zero followers, two or three commits.
  2. The program isn't built from the code. The README tells you to run an .exe or a zip from the releases, but nothing in the repo produces it.
  3. The download lives somewhere else. A button to Dropbox, a file host or an unknown site, often with an archive password.
  4. It asks for too much. "Run as Administrator", "disable your antivirus", or a one-line PowerShell command to paste.
  5. The README is empty and the bait is elsewhere. Check CONTRIBUTING.md too.
  6. The name belongs to someone else. A real project's name with a prefix or suffix added, and none of its code.
  7. The stars came in steps. A tool like a star-history chart shows big jumps on a few days instead of a steady curve.

Where this leaves me

A ranking that rewards growth will always attract people who sell growth. I don't think I'll ever be "done" with this, any more than a spam filter is done with spam. Every week the tricks change a little: a different zip size, a few more code files, a real project's name.

What I can do is make each trick cost more, and show the evidence. Upstar's rankings are free, with no account: upstar.dev. If you follow a topic there and see something that looks off, tell me. And if you've seen these star deliveries or this "resp" family somewhere else, I'd love to compare notes in the comments.