Like Audit: spot farmed likes on Hugging Face models before you trust a trending repo

I found no product, extension or API that scores whether a Hugging Face model's likes are real. The nearest neighbors are GitHub fake-star research (StarScout, the archived Astronomer), enterprise model scoring that takes raw likes at face value (Endor Labs), and file-level malware scanners shown on model pages (JFrog, Protect AI). The May 2026 Open-OSS incident gives the clearest timing hook: bot likes pushed a malware repo to number one on trending. A trust badge that flags burst-created, inactive or pattern-named likers would have caught it. The biggest risk is not a startup rival. It is Hugging Face itself, which can see the raw data, has every reason to clean its trending list, and could quietly discount farmed likes. Security vendors like HiddenLayer, Socket or JFrog could also fold this into their existing scanners. Individual ML engineers are unlikely to pay for a browser extension. More promising buyers are security and procurement teams, who could get the scores as a feed added to supply-chain tools, and investors checking traction claims. Fake downloads are harder to spot from outside the platform, which could be a lasting edge. The cleaned trending newsletter is cheap to launch and good for distribution, but it will make little money.

• HiddenLayer (Open-OSS privacy-filter research)

  In May 2026, HiddenLayer found that a fake OpenAI repo pushed to number one on Hugging Face trending had 657 of its 667 likes from bot-named accounts. It is one-off threat research, not a product, but it shows the problem is real and that security vendors are watching it. HiddenLayer sells a model scanner and could add like analysis.

• Endor Labs Hugging Face model scores

  Scores Hugging Face models on activity, popularity, security and quality for enterprise buyers. Its popularity score uses raw likes and downloads and does not check whether they are real. It is the closest commercial neighbor and could plausibly add an authenticity layer.

• JFrog / Protect AI scanners on Hugging Face

  Malware and security scans shown directly on Hugging Face model pages. They check model files, not engagement. They own the trust real estate on the page, and that is a distribution threat.

• StarScout (CMU/Socket/NCSU)

  A research tool that found about 6 million suspected fake GitHub stars using low-activity and lockstep patterns. It is GitHub-only, but its methods carry straight over to Hugging Face likes, and Socket could productize it across ecosystems.

• Astronomer

  An open-source tool that scores whether a GitHub repo's stargazers are real. It is archived and has not been updated since 2020. It is a GitHub analogue of the trust badge idea and shows the concept but no current competitor.

• Model Trust Scorecard

  An open-source project that checks benchmark claims in model cards against leaderboards. It is a trust score for claims, not for likes, so it complements rather than competes.

• HuggingFace Checker (Chrome extension)

  An existing Chrome extension for Hugging Face pages. I could not confirm what it does, and nothing suggests it analyzes likes. Worth checking by hand.

Checked 2026-10-18