Five independent actions in seven days. This week, Codeberg banned AI-generated code in a 358-144 member vote (July 22). Debian opened a formal General Resolution to potentially ban LLM contributions across the entire project (July 24). And today, GitHub restructured its bug bounty program to combat AI report flooding, cutting top public payouts by roughly two-thirds. This follows curl ending its six-year bounty program in January and Google halting AI-generated CVE reports in March. Open source is not gradually growing skeptical of AI contributions. It is formalizing rejection — and the pace is accelerating.
The Economics Broke First
Every ban and restructuring this week traces back to the same underlying economics. Generating a plausible-sounding vulnerability report now costs pennies in tokens. Evaluating whether it describes a real security flaw still costs an expert an hour or more. That asymmetry is not theoretical — it has already collapsed multiple programs.
curl ran its bug bounty for six years, paying out over $100,000 on 87 confirmed vulnerabilities. Then AI-generated reports flooded in, and the valid-report rate fell from one-in-six to one-in-twenty or worse. Creator Daniel Stenberg’s team was spending more time debunking garbage than writing code. curl shut the program down in January 2026. GitHub’s situation mirrors this: today’s restructuring cuts the maximum public payout to $10,000 for critical findings, reserving higher rewards for an invite-only VIP tier that requires demonstrated track records. Screenly found the same pattern — 331 reports in six months, only 39 actual vulnerabilities. Meanwhile, the EU Cyber Resilience Act, effective September 2026, is about to legally mandate vulnerability disclosure programs for companies selling into Europe. The open-source community is being pushed to maintain security infrastructure that AI-generated noise is actively destroying.
Related: GitHub Bug Bounty Cuts: AI Report Spam Forces Two-Tier Program
A Week of Formal AI Bans in Open Source
Codeberg’s July 22 vote targeted AI-generated “ghost projects” consuming platform resources — storage costs at the donation-funded forge have ballooned from roughly €700 to €3,700 per SSD, a fivefold increase that cannot be sustained on volunteer funding. The amended Terms of Use now prohibit projects “mostly consisting of code written by generative AI tools,” specifically naming Claude and OpenAI Codex. The vote passed 358 to 144 on roughly 50% member turnout — a clear majority, not a fringe decision.
Debian’s stakes are higher. The Debian General Resolution, now in discussion phase, would ban LLM contributions from source packages, packaging work, official documentation, translations, and project communications. A competing proposal would instead permit AI-assisted work with mandatory “Assisted-By” or “Generated-By” disclosure tags, placing full accountability on contributors. Debian hasn’t voted yet — but the fact that the GR exists at all matters. Debian underpins Ubuntu, Mint, and hundreds of downstream distributions. Whatever Debian decides on this will echo through the GNU/Linux ecosystem for years. Zig, NetBSD, GIMP, and QEMU had already implemented outright bans before this week. The movement is no longer fringe — it is reaching infrastructure that billions of devices run on.
The Enforcement Problem Nobody Is Solving
Here is the problem no one in the “ban AI contributions” camp has solved: there is no reliable technical mechanism to detect AI-generated code. Codeberg explicitly acknowledged its “limited resources and overburdened workforce.” In practice, enforcement means projects that draw complaints get flagged — everyone else continues quietly. Critics, including Flask creator Armin Ronacher, have called this counterproductive. You can vote to ban AI contributions. You cannot reliably detect them at scale.
The Linux kernel took the opposite approach. Rather than banning AI-assisted code, the kernel project requires an “Assisted-By” tag — human contributors remain accountable, but the use of AI tools is disclosed rather than hidden. Zig went further in the other direction: its “Contributor Poker” philosophy, articulated by VP Loris Cro, holds that maintainers invest review time in the contributor as a human with potential, not just in the code submitted. AI breaks that signal entirely. Both positions are coherent. However, the kernel’s approach is actually enforceable. Symbolic bans with no detection mechanism may do more to frustrate legitimate developers who use AI responsibly than they do to stop bad-faith flooding — the people actually spamming pull requests and CVE databases are not going to read the Terms of Use and comply.
What Developers Using AI Tools Need to Do Now
The practical impact of this week’s actions is specific and navigable. Most open-source projects have not banned AI contributions. However, the rules are no longer uniform, and assuming they are is how you waste time and damage relationships with maintainers.
Before contributing to any project, check its contribution guidelines for an explicit AI policy. Projects like Zig prohibit LLM assistance “in any form,” including IDE-integrated tools — submitting a Copilot-assisted PR is a violation whether you declare it or not. For bug bounty programs, the AI report flood has fundamentally changed triage economics across the industry. GitHub’s new HackerOne signal requirement means new researchers get four initial submissions to establish credibility before limits apply. Reports that don’t demonstrate real security impact will now accelerate you toward submission limits faster. The practical guidance hasn’t changed: contribute work you genuinely understand, start small, and engage meaningfully in code reviews. AI can help you understand a codebase — but submitting work you cannot explain or defend is a fast path to being filtered out by the increasingly automated gatekeeping these bans are creating.
Key Takeaways
- Five major open-source platforms and programs took formal action against AI contributions in a single week — Codeberg, Debian, GitHub, curl, and Google. This is a systemic shift, not isolated incidents.
- The core driver is economic asymmetry: AI reports cost pennies to generate but hours to evaluate. Bug bounty programs are collapsing under this load just as EU regulation makes them legally mandatory.
- These bans largely lack enforcement mechanisms. There is no reliable technical tool to detect AI-generated code at scale. Enforcement is complaint-driven and self-reported.
- Check contribution guidelines before submitting to any project. Zig and similar projects prohibit all LLM assistance; most projects do not — yet. Rules are diverging rapidly.
- The Linux kernel’s “Assisted-By” tag model is the most enforceable compromise. If Debian adopts a similar approach rather than an outright ban, it may become the de facto standard for the broader ecosystem.
