curl Is Drowning in AI-Assisted Security Reports, and the Cost Is Human

On May 26, 2026, Daniel Stenberg, the founder and lead maintainer of curl, published a post titled “The Pressure.” It is not a manifesto. It reads like a status report from someone who is tired. He is now receiving more than one security report per day. That is four to five times the volume he saw in 2024, and twice what he saw in 2025. Much of it is AI-assisted, the kind of report that looks plausible, cites a real function, and falls apart the moment a human who understands the codebase reads it carefully.

Here is the part that should stop you. While the volume climbed four to five times over, the real severity stayed flat. curl has 12 confirmed CVEs pending and projects roughly 30 for all of 2026. The last report rated HIGH was in October 2023. So the signal has not grown. Only the noise has.

This is the demand-side break in open source, and curl is the clearest primary source we have for it.

Volume Up, Severity Flat: The Shape of the Problem

We have written before about the two-front crisis hitting the software supply chain, where AI pressures open source from both the contribution side and the consumption side. That essay made the argument in the abstract. Stenberg’s post is the argument made flesh, with dates and counts attached.

Look at the asymmetry directly. Reports per day: more than one. Year-over-year growth: four to five times since 2024. Confirmed serious findings: flat. When the cost of producing a security report drops toward zero, the volume of reports rises toward infinity, and the fraction of them that matter falls toward zero. The economics are not subtle. A language model can draft a CVE-shaped report in seconds. It cannot, on its own, tell whether the report is real. That verification still costs a human being a full read of the relevant code.

So the maintainer absorbs the entire delta. Every low-quality report still has to be opened, read, reasoned about, and refuted, often with a written explanation, because the reporter may be acting in good faith and may also be a real researcher having a bad day. You cannot auto-reject your way out of this without eventually rejecting the one report in thirty that would have mattered.

curl is not a side project. It ships in roughly 30 billion installations. It is a 30-year-old codebase that sits underneath cars, phones, payment systems, and most of the internet’s plumbing. The person reading these reports is working 50-plus-hour weeks, seven days a week. The substrate the world runs on is being defended by someone the volume curve is actively trying to bury.

This Is Not a Triage Problem

The reflexive response is to reach for better filtering. Add a CAPTCHA. Require a proof of concept. Charge a deposit. Train a classifier to score incoming reports. Stenberg has tried versions of this, and the bounty programs have their own rules. Filtering helps at the margin. It does not touch the root.

The reason filtering fails as the answer is structural. Every filter you add to reject slop also adds a step that a legitimate, nervous, first-time reporter has to clear. Raise the bar high enough to stop the flood and you also stop the responsible-disclosure pipeline that security depends on. The maintainer is stuck optimizing a single dial between two failure modes: drown in noise, or silence the signal. There is no setting on that dial that solves the problem, because the problem is not the dial.

The problem is that a single human is the only verification layer, and that human is unpaid or underpaid relative to the value they protect, and there is no organization standing behind them. When a corporate security team gets buried, it hires. When curl gets buried, Stenberg works the weekend.

That is the real exposure, and it is not technical. It is the absence of an institution.

The Missing Life-Boat

Some open-source projects sit under an umbrella. The Cloud Native Computing Foundation, the Apache Software Foundation, the Linux Foundation: these give a project legal cover, funding channels, and crucially a way to convert demand pressure into hired capacity. Independent projects like curl have none of that by default. There is no life-boat. When the wave comes, the maintainer swims.

This connects to a theme we have traced in software slop and the governance of attention: AI does not just generate low-value output, it redirects scarce human attention toward sorting that output. Every hour Stenberg spends refuting an AI-drafted non-vulnerability is an hour not spent fixing a real one, mentoring a contributor, or resting. The cost is not abstract. It is measured in burnout, and burnout in a single-maintainer project is a supply-chain risk for everyone downstream. When the maintainer stops, 30 billion installations inherit the silence.

The uncomfortable truth for the companies that depend on curl is that they have outsourced a load-bearing function to a volunteer and never put it on a contract. They run curl in production. They ship it to customers. They have never sent a dollar or an engineer-hour to the person who keeps it safe. That arrangement worked when the volume was manageable. AI just made the volume unmanageable, and the bill is now visible.

What Governance Actually Means Here

If you lead a team that builds on open source, and almost every team does, the governance question is not “how do we triage AI-generated reports.” It is “have we funded the humans our stack depends on.” Those are different questions with different owners. The first lands on a security engineer. The second lands on a budget holder.

The fix that scales is commercial. Funding. Support contracts. Paid maintainer time. An institution willing to convert a recurring line item into hired verification capacity, so that a flood of reports meets a team instead of one exhausted person. curl already offers commercial support through wolfSSL, which is exactly the mechanism this calls for. The companies shipping 30 billion installations could fund that capacity many times over and would not notice the cost on their balance sheets.

This is the same lesson the supply-chain two-front crisis pointed at, now with a name and a face. Governance of AI noise is not only a filtering discipline. It is a funding discipline. You cannot automate your way out of a problem whose only real solution is paying a human to keep verifying.

Do This Now

Three moves, in order of who owns them.

Inventory your dependence. List the open-source projects your product cannot ship without. For each one, answer a single question: is it backed by an institution, or by a person? If the answer is a person, you have a concentration risk that no contract currently covers.

Fund the load-bearing ones. Pick the projects in that list with no umbrella organization. Set up sponsorship, a support contract, or paid maintainer time this quarter. For curl specifically, commercial support exists through wolfSSL. The cost is trivial next to the cost of the maintainer walking away.

Govern your own reporting. If your security team files reports upstream, make sure a human verifies every one before it goes out. Do not be the source of someone else’s flood. The discipline you want others to have starts with the reports you send.

The teams that stay safe over the next two years are not the ones with the best triage filters. They are the ones who treated the humans under their stack as infrastructure, and paid for them like infrastructure, before the wave arrived.

This analysis synthesizes The Pressure (Daniel Stenberg, May 2026).

Victorino Group helps teams govern the AI-generated noise flooding their security and review pipelines. Let’s talk.