A Click-Fraud Bot Showed Me 5,000 GitHub Accounts Anyone Could Try to Reset

It started with a boring campaign
Recently I started doing malware research on GitHub, and I found something that looked boring at first glance.
38 repositories, on 20 old accounts, all with the same new GitHub Actions workflow. Every two hours it opened a browser on GitHub's free runners and clicked paid links. Ad fraud on someone else's compute. Nothing you'd write a post about.
But the accounts were strange. They were created between 2008 and 2017. Their last commits were from 2015, 2016, 2017. Then nothing for years, and in 2026 a workflow shows up.
Dormant accounts don't wake up on their own. Somebody was signing in to them. I wanted to know how.

The pattern in the commits
So I stopped looking at what the bot does and started looking at who committed it.
Every git commit carries an author email. I compared the email on the new 2026 commit with the email on the account's original commits from a decade ago. In 11 of the 19 accounts I compared, it was the same address.
And it was a Mailinator address. Or Yopmail. Or Guerrilla Mail.
If you never used these: they're public inboxes. No signup, no password. You type a name and you read whatever mail arrived for it. Anyone can do that for any name.
That was the moment it clicked. The operator wasn't picking random old accounts. It was picking accounts whose email anyone can read.

Testing what the attacker seemed to know
A pattern is a hint, not proof. So I tested it on myself.
I have an old GitHub test account with a Mailinator address. I asked GitHub to reset its password. The reset email landed in the public inbox. I opened the link, set a new password and signed in.
Nothing to crack, nobody to phish. And you don't have to guess which inbox to open, because the address is printed in the account's public commit history.
GitHub blocks these addresses at signup today. Old accounts kept them.
Turning their pattern into my search
Once I knew what the operator was selecting for, I could run the same selection across all of GitHub.
On October 14, 2026, I found 5,552 live accounts whose old commits used a public-inbox address. 917 of them were already running the operator's workflow.
The campaign was much bigger than the 38 repositories I started with: 1,014 repositories owned by 965 accounts. In 836 of the 1,007 still online, the first 2026 commit was made in GitHub's web editor under a public-inbox address. You can only do that when you're signed in and the address is verified on the account.
I can't see how the operator signs in. A reset through the public inbox fits. Buying old accounts fits too. Only GitHub's logs can tell. Either way, its pattern led me to a door that I could open myself.
Then the operator confirmed it for me
I reported everything to GitHub and Hugging Face on October 14, 2026. The report included 860 more accounts that still had a public-inbox address linked and were not part of the campaign yet.
Four days later, 18 accounts from that list added the campaign's workflow. The first at 14:12 UTC, the last at 14:49. None had committed anything since early 2017. Every commit was made in the name of the account's old public-inbox address.

I had predicted which accounts were next, and the operator went and took them. Four more from the list had been running the workflow since September 12 and I had missed them.
On November 12 I checked again. 21 of the 22 repositories were still public, and every one had a workflow dispatch that same day.
The bot doesn't keep its scripts on GitHub. It downloads them from Hugging Face datasets on every run. I found 23 of them, spread over seven accounts: gfdg34fsd, samilod9, sjkhfuk, sjkhfuk1, swelam100, swe1000 and hiodas94.
In August the bot picked up a second job. On August 20, 2026, a new dataset called gfdg34fsd/newe kept the ad clicker and added a crypto miner (SRBMiner, mining Xelis and Zephyr coins).
Hugging Face removed the datasets I reported. The operator opened a new one within hours. On November 12 that one, sjkhfuk1/foffo, was still up: 442 repositories were fetching from it, and its 30-day counter showed 993,931 downloads. Those are machines coming back for the script, not people.
What GitHub said
On November 12, GitHub closed my report as Informative. They confirmed the behavior and called it a known issue without significant security risk.
I get the logic. These are abandoned accounts and the operator is clicking ads. I looked for worse and didn't find it: no campaign commit in a repo with more than 110 stars, and none of the npm accounts I found on public inboxes belongs to the campaign.
But the operator had write access to every one of those repositories, and it chose ads. The door doesn't care what you carry through it.
Attackers already did the research
Here's what I took from this. I would never have gone looking for "GitHub accounts with a Mailinator address from 2016". It's too odd a thing to think of.
The operator thought of it, and then left the evidence in public. The email on a commit. The age of an account. The exact minute 18 accounts woke up. A low-effort ad bot turned out to be a map to a real gap in how old accounts are protected.
So when I see a dull campaign now, I ask a different question. Not "what is it doing?" but "why these accounts?"
Your turn
Check your own old accounts. If a GitHub, npm or PyPI account of yours still has a throwaway address on it, replace it and turn on two-factor authentication. Then look at your old repos for a workflow you didn't write.
If you run a platform, don't let people register with public-inbox addresses, and don't send password resets to the ones already on file. Blocking them at signup only protects new accounts. The old ones are still open.
Have you ever learned something real from watching what an attacker picked? Tell me in the comments.
How I counted: the 5,552 accounts come from GH Archive push events for 2014 to 2019, counted on October 14, 2026. Everything else is public GitHub and Hugging Face metadata. I sent no reset requests to anyone's account but my own. Accounts tied to disposable inboxes were described as a risk in earlier research.