In January, the curl project — the tiny, ubiquitous tool that moves data for roughly every connected device on Earth — announced it was done paying people to find its bugs. The bug bounty program, which had run for years through HackerOne, would close at the end of the month. The project’s own policy file now reads like a man who has been pushed past politeness: “Up until the end of January 2026 there was a curl bug bounty. It is no more.” A later revision is blunter about why: a bounty “gives people too strong incentives to find and make up ‘problems’ in bad faith that cause overload and abuse.”
The details are almost comic. Daniel Stenberg, curl’s founder, reported receiving seven HackerOne submissions in a single sixteen-hour stretch, and around twenty in the opening weeks of 2026 — none of which identified an actual vulnerability. He told Axios that fewer than five percent of the reports submitted in 2025 were legitimate, and described “the never-ending slop submissions” as taking “a serious mental toll on the curl developers.” The project’s security.txt now warns that people submitting “crap” reports “will be banned and ridiculed publicly,” which is not the tone of a thriving intake pipeline. It is the tone of a shopkeeper nailing plywood over the door.
The easy reading of this story is aesthetic: AI-generated reports are bad, the prose is greasy, the findings are hollow. But that misses what actually broke. The reports were always going to be mostly wrong — that was true of human submissions too. What changed is the cost of producing one. A plausible-looking vulnerability report, complete with reproduction steps and a suggested severity, now costs its author a prompt. Evaluating it still costs a maintainer real hours. Sam Robinson of the Open Source Security Foundation told Axios that a popular project might once have fielded two or three bug reports a week, and that some now receive hundreds at once, each taking two to eight hours to validate. Whether or not his numbers generalize, the shape of the problem does: generation has become nearly free, while verification remains stubbornly expensive, and the difference is paid in human attention by whoever was unwise enough to leave a door open.
That asymmetry is the actual story, and it reaches well past security teams. Every open venue on the internet — every journal with a submissions inbox, every marketplace, every forum, every open-source project with a CONTRIBUTING.md — was built on an assumption so deep it was never stated: that contributing took effort. Effort was the spam filter. Writing a coherent bug report, a short story, a pull request, even a thoughtful forum post required enough investment that the mere existence of the submission was weak evidence of sincerity. Moderation, peer review, and maintainer triage were all calibrated to a world where the supply of contributions was limited by the number of hours humans were willing to spend. Nobody designed these systems. They grew in soil that happened to contain friction, and they mistook the soil for a law of nature.
Generative tools salted that soil. And the venues are discovering that they have exactly two options, both bad. They can add friction — applications, invitations, credentials, fees, waiting periods — which filters out the slop and also, not incidentally, the unknown newcomer who was the entire point of being open. Or they can stay open and drown, which is a slower way of achieving the same closure.
You can watch projects pick their poison in real time. Ghostty, the terminal emulator, is moving to an invitation-only contribution model — open source in license, members-only in practice. GitHub, in a blog post pointedly titled “The Eternal September of open source,” says it is exploring criteria-based gating, like requiring a linked issue before a pull request, and automated triage of contributions against a repository’s own guidelines; it has already shipped controls letting maintainers restrict pull requests to collaborators or turn them off entirely. Stenberg chose differently, attacking the incentive rather than the door: “By removing the money incentive, we remove the incentives for submitting made-up lies in the hope of getting a bag of money.” Kill the bounty, keep the inbox. It is a coherent bet, but note what it concedes — the problem was never the prose quality of the reports. It was that someone could produce a lifetime supply of them before lunch.
The historical irony is that the old-timers saw this coming, sort of. The Linux kernel introduced its Developer Certificate of Origin in 2004, and its SubmittingPatches guide has always been a small obstacle course of formatting rules and sign-offs. For years this read as cranky gatekeeping, the priesthood defending the temple. Now it reads as load-bearing infrastructure: friction installed deliberately, by people who understood that a contribution you had to work for was a contribution you could afford to read. What’s new is that this friction used to be optional. The internet’s openness was subsidized by scarcity, and the subsidy has been withdrawn.
There is a temptation to frame all of this as a story about AI, but AI is just the solvent. The deeper revelation is about what “open” ever meant. For thirty years, openness was cheap to offer because almost nobody could afford to take you up on it at scale. The open door worked because walking through it cost something. Now that walking through costs nothing, every institution built on the door has to decide what it actually valued: the openness, or the scarcity that made the openness survivable.
curl’s answer is on the record, in its own repository, written in the flat voice of a man updating a policy file at the end of a long losing argument. The bounty is no more. The reports still come. Somewhere in that sentence is the future of every open venue on the internet: not closed, exactly — just tired, and quietly relearning that a door is only as good as whatever it costs to knock.