deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

Star-buying networks push fake, malware-laced GitHub repos into trending lists

A dev.to writeup by the developer of GitHub-ranking site Upstar documents stars delivered in lockstep to prop up fake repos, including campaigns that bait users into running malware.

Star-buying networks push fake, malware-laced GitHub repos into trending lists

The developer behind Upstar, a site that ranks fast-growing GitHub repositories, says coordinated campaigns are purchasing stars in batches to push fake projects toward trending visibility. According to a detailed post on dev.to, one such repo, "google-ultra", which promised paid Google AI tiers for free, reached the number-one Python spot on Upstar before the author blocked it and GitHub removed it within a day.

Upstar is a solo project run from Abidjan that scores repositories across 69 topics every hour, favoring fast relative growth over absolute star counts. That design is exactly what purchased stars exploit: a few hundred cheap stars on a brand-new repo can lift it into the top of a niche topic. The site's filters, which combine rules, a trained classifier and LLM judges, have blocked more than 3,800 repos since October.

Stars delivered in lockstep

The clearest evidence came from stacking day-by-day star counts. Forty repositories, each owned by a different account with zero followers, all received exactly 8 stars on November 14 and exactly 17 stars on November 17. Twenty-five of them had also gained 3 to 5 stars on November 12, and 38 of the 40 owning accounts were created on October 27 or later, many in clusters.

The author flags an important caveat: seven legitimate projects, among them audiobookshelf, newsnow and mlx-audio, happened to hit the same star counts on those days. Their curves show steady daily gains rather than sudden steps, which is why identical numbers alone prove nothing — the account history is what turns coincidence into a pattern.

Real names, empty repos

Some fakes borrow the name and description of a genuine open-source project and add a prefix. One, "resp-expandModeIntegerScaler", mimicked a real Godot addon. Inside, there was almost nothing: an HTML page and a CONTRIBUTING.md. With the README blank, the hook sat in the contributing file — an off-site download button, a password for the archive, and an instruction to run as Administrator.

The "resp" prefix appears in repo names tied to 136 young accounts, and 66 further accounts were created within a 35-minute window on October 28, many with names built from the same few words.

A cheat network assembled in 47 seconds

A second campaign sells cheats for a game called Rob Run. Six accounts were created in a 17-minute span on October 14, five of them within 47 seconds of each other, each publishing its repo roughly two minutes later. Every README tells users to run robrun-hk.exe — a file the repo's code does not build, delivered instead as a hand-uploaded zip in the releases section.

The operators dressed things up carefully, according to the post: each zip had a distinct size between 1.65 and 2.92 MB, and the repos carried 9 to 16 plausible-looking code files. Even forks appear purchased — 41 of one repo's 80 forks came from accounts all created on November 7. One variant used a PowerShell one-liner to fetch a script from GitHub Pages that decoded a payload with base64 and XOR; the author did not analyze what it installs.

What the filters missed

The post is candid about detection gaps. A rule flagging "run this exe" prompts only fired on near-empty repos, so the padded Rob Run repos slipped through. A duplicate-file rule required exact matches, which differently sized zips defeated. LLM judges saw real code and a GitHub-hosted zip and let the cheats stay, while older approvals were never re-read after policies changed. Most strikingly, a classifier scored the Rob Run repos as high as 0.998 out of 1 — but it could only confirm blocks a judge had already proposed, so its score went unused when the judge said keep.

Those gaps are being closed: any repo the classifier scores at 0.9 or above now escalates to a stronger model, and since November 18 the classifier alone has blocked 268 lures, with no blocks reversed so far.

How to spot one yourself

The author's checklist: an account as young as its repo with zero followers; an executable the code never builds; downloads hosted on Dropbox or unknown sites, often password-protected; demands for admin rights or disabled antivirus; an empty README with the bait in CONTRIBUTING.md; a real project's name with a prefix attached; and star-history charts showing big jumps on a few days instead of steady growth. No single sign is proof; two or three together should end the visit.

Why it matters

Star counts work as trust signals. They shape what developers discover, adopt and — in these campaigns — execute on their machines. When stars can be bought in identical batches, any ranking that rewards growth becomes attack surface, and the payload is aimed squarely at people who trust GitHub enough to paste a PowerShell command. The takeaway for developers, ranking sites and platform trust-and-safety teams alike: growth numbers need corroboration, and a repo's account history, its age, and whether its code actually builds the artifact it asks you to run matter more than its star total.

  • #github
  • #security
  • #malware
  • #open-source

Related posts