Back to Blog

How to protect a free for open source program

A thin repository can take a year of free builds from a real maintainer. Why free for open source program abuse happens, and which rules hold up.

The repository on the application has an MIT license, a two-line README and one commit. That commit landed 11 minutes before the form was submitted.

You've opened it in month five of your open source program, because the build logs looked odd. The reviewer's note is still there: "Looks fine, welcome aboard." It did look fine. Then you scroll down the approved list and find 22 more just like it.

That list is a scenario, and so are the numbers below, so swap in your own. Say you run a hosted CI service. An approved project gets your Team plan free for 12 months: 25,000 build minutes a month, which lists at $199. Assume a build minute costs you $0.004 in compute. Here's free for open source program abuse in arithmetic:

  • One project at full use: 25,000 × $0.004 = $100 a month
  • One 12-month grant: $100 × 12 = $1,200 at cost, or $199 × 12 = $2,388 at list
  • A fully used 14-day trial of the same plan: $100 × 14 ÷ 30 = about $46.67
  • One bad approval, in bad trials: $1,200 ÷ $46.67 = about 26

Open the build history for those 23 accounts. The public repository built once, around the day of the application. Every build since came from private repositories on the same account, because your grant attached to the account. That's somebody's commercial software compiling on your open source budget. When you write to them, one small agency cheerfully admits to nine of the approvals: 9 × 25,000 = 225,000 free build minutes a month, worth $1,791 a month at list (9 × $199).

Say all 23 were approved in the program's first month:

  • 23 approvals × $100 = $2,300 a month
  • $2,300 × 5 months = $11,500 already spent
  • $2,300 × 12 months = $27,600 if nothing changes
Month 123 approvedMonth 3$6,900 spentMonth 5$11,500, you find itMonth 9$20,700 spentMonth 12$27,600 spent
Month 123 approvedMonth 3$6,900 spentMonth 5$11,500, you find itMonth 9$20,700 spentMonth 12$27,600 spent

The $11,500 is gone. You can end a grant today, but you can't take back five months of finished builds, and there's no invoice to collect on. If your company spends $40,000 a month, the full $27,600 is about three weeks of runway ($27,600 ÷ $40,000 = 0.69 months).

The number that really stings is 23 out of 60. You capped the program at $6,000 a month, which funds 60 fully used projects ($6,000 ÷ $100), so thin repositories hold 38% of the places. Your waitlist has 31 names on it. One is a date-parsing library that thousands of other projects install, kept alive by someone working evenings. They applied in April. Last week they withdrew, with one line: "No worries, we moved the tests to an old laptop under the desk." Your program exists for them, and a two-line README got their place.

Paying customers got less too. Those accounts share your runner fleet, and 23 × 25,000 is 575,000 build minutes a month. If most of that lands in weekday working hours (22 days × 480 minutes = 10,560 minutes), that's about 54 runners busy at once (575,000 ÷ 10,560). On Monday at 10:40 a.m., a customer on your $199 plan waits 19 minutes for a two-minute hotfix build to start, then asks for a credit.

Now you have to decide. Revoke all 23 and one of them could be a real project that started small, whose maintainer posts that your company turned on open source. Leave them alone and the waitlist keeps waiting. Close the program and you lose the thing maintainers recommended you for, while the announcement is still pinned to your blog.

Why free for open source programs attract thin repositories

A thin repository exists to pass your form. It has the files you asked for and not much else, and your form can't see what it was made for.

Programs like this are built on generosity, which is the right instinct. Someone reads a public page for a few minutes and hands over a year of a paid plan. A normal trial gives away 14 days and ends by itself. This gives away 12 months on the strength of something that costs nothing to create, and because the grant lands on the account, it covers whatever that account builds next. Startup credit programs get hit the same way.

The obvious fix is to raise the bar. Require 1,000 stars and the two-line READMEs disappear, along with most of your waitlist. That date library has 140 stars, because people install it and never visit its page. Every real project also had a first month when it looked exactly like a thin repository. A high bar mostly turns away the small maintainers you started this for.

It also answers only one of your two questions. The first is whether this project is real open source. The second is whether this application is a repeat, an eighth or ninth grant under another name. They need different answers.

The basic rules for open source plans, and how each one breaks

Most programs try these, and each one does catch something.

Basic ruleWhat it catchesWhat it missesWho it wrongly blocksWhat it costs to keep working
License file requiredProprietary codeAnyone who adds a license fileAlmost nobodyA list of licenses you accept
Star minimumBrand-new repositoriesAnything over the numberUseful libraries with few starsArguments about the number
Minimum repository ageRepositories made for the formAnyone willing to waitNew projects with real users"Come back later" emails
Email verificationTypos and dead addressesAny real second inboxMaintainers whose code lands in spamDeliverability work
Disposable-domain blocksKnown throwaway servicesNew domains and ordinary mailboxesPeople using privacy forwardingA list someone updates forever
Gmail dots and plus addressingOne inbox written several waysA second real mailboxPeople who share one inboxUpkeep as mail services change
CAPTCHA on the formSimple scriptsPeople applying by handPeople using screen readersFees and lost applications
Phone or card requiredCasual repeat applicationsAnyone with another number or cardVolunteers without a business cardText charges and fewer applications
One per IP addressRepeat forms from one connectionAny other networkCo-maintainers on one campusExceptions and tickets
VPN blocksKnown VPN rangesEverything off your listContributors who want privacyAnother list to maintain
Shorter grantsValue per bad approvalThe next applicationMaintainers who have to reapplyRenewals to process
Manual reviewWhat a person can spotWhatever looks fine for 12 minutesWhoever looks odd to a tired reviewerHours every week

The email rows come down to how email works. Gmail ignores dots, and plus-tagged mail reaches the same inbox, so one mailbox can look like many applicants. A CAPTCHA has limits of its own.

Repository rules break in a more awkward way. Say you add a six-month age rule this week. A 40-day-old library with real users gets a "come back in five months" email, and anyone willing to wait applies in five months. You've delayed the abuse and insulted a maintainer.

So it all lands on the reviewer. At 20 applications a week and 12 minutes each, that's 4 hours a week, or 208 hours a year. Nothing on the page they're reading says whether the person filling out the form already holds eight grants.

Program rules, illustrativeWorks but costs goodwillWorth havingCheap and weakHurts and barely helpsEasy on maintainersHard on maintainersStops few thin reposStops many thin reposLicense file ruleRepo age minimumShort renewalsPublic repos onlyManual reviewStar minimum
Program rules, illustrativeWorks butcosts goodwillWorth havingCheap andweakHurts andbarely helpsEasy on maintainersHard on maintainersStops few thinreposStops many thinreposLicense file ruleRepo age minimumShort renewalsPublic repos onlyManual reviewStar minimum

How to limit what a bad approval can cost

Do these first. None of them catches anyone, but each one changes what a thin repository can win.

Attach the grant to the repository, public code only. Private builds on that account bill as normal. That removes the biggest reason to bring a thin repository at all.

Publish what counts. Say which licenses you accept, that development has to happen in public, and that a commercial product doesn't qualify.

Grant 90 days first, then renew automatically while the public code keeps building. Real maintainers do nothing. A bad approval now costs at most $100 × 3 = $300 instead of $1,200, and 23 of them cost $6,900 instead of $27,600.

Give sponsored builds their own queue and a hard monthly cap, so a bad month can't reach paying customers.

Those steps shrink each bad approval to a quarter of its old cost. They do nothing about the number of approvals, and the agency's tenth application arrives with a new repository and a new inbox that pass every row in that table. What's still open is whether that claim should go through at all.

Evaluate each application before the grant goes out

You don't have to build your own abuse prevention system to answer that. AbuseZero takes one API call, made right before the grant goes out, and returns a clear allow or deny under rules you control. Where your domain list goes stale, it checks mailbox aliases, disposable domains and mail routing. Where a one-per-IP rule lands on a whole campus, it sees beyond an IP address with VPN, hosting, network reputation, Tor and proxy intelligence. You get multi-layer AI risk analysis that examines 140,000+ signals, detects linked claims and flags abuse that single checks miss, and single checks are all that table ever had. You also get a proprietary algorithm, with AI that checks the claim again. AbuseZero recognizes repeat devices across new accounts, even when email addresses keep changing. Creating a fresh account isn't enough to erase previous activity.

Your form still decides what counts as open source, and a person should still read the project. Set the grant's eligibility, then start in Observe mode, which shows you what would be denied before anyone is blocked. Maintainers you already trust can go on an allow list. Send every deny to a reviewer instead of an automatic rejection, because turning away a real maintainer costs you more than a wrongly denied trial, and if the reviewer approves it you can record the grant after a deny. The quickstart protects your first benefit.

Frequently asked questions

Should an open source plan cover private repositories?

Usually not. Covering public code keeps the grant tied to the work you meant to sponsor. If maintainers need somewhere private for security fixes, give them a small separate allowance.

Is it abuse if one maintainer applies for several projects?

No. Plenty of maintainers run several real projects. Decide how many one maintainer can hold, write it down and read each project. A one-account-per-person policy needs the same care.

What should I do about thin repositories I've already approved?

Look at where each account's build minutes went before you write to anyone. Then say what you saw, give them 14 days to reply and end the grant at its next renewal if nothing changes. Some will be real projects that started small, so word the message carefully.

Give the next grant to a real project

Keep reading

Copying your prompt…

Copied the to your clipboard. Give it to your coding agent & it will handle the rest.