Lots of people who are against the use of AI for all its ethical and inherent issues tell me this. Just ignore it. Don’t let big AI control us.
The “slop” they refer to is anything created that involved an AI, and in our case (the curl project) it is in particular vulnerability reports.
“Ignoring the slop” cannot be done and it doesn’t help. Vulnerability reports created with the help of AIs point out actual problems, and sometimes vulnerabilities. Ignoring real vulnerabilities because they were found by AI does not make users safe and secure. It makes them vulnerable. We don’t want that.
I personally don’t refer to this as AI slop because the reports are typically accurate and correct, even if much too wordy. I think “slop” implies that it is wrong, misleading or otherwise bad content.
One can certainly object to the use of AI, but when we get real vulnerabilities reported and proven, we cannot in good faith pretend they don’t exist just because we think the tech involved in digging them up is questionable. That would be a disservice to, or even betrayal of, our users.
We make curl secure for our users.
Obligations
An Open Source project typically gives no guarantees at all, so there are no legal requirements that force us to fix flaws.
We are however a project with high goals of quality and stability and we would jeopardize our name and track record if we now would stop delivering on that. Also, a few of us involved in the project also sell curl related support and services, where at least part of that means providing secure curl releases, so we still have this contractual obligation to a set of customers. And they wouldn’t be customers in the fist place if the Open Source project didn’t already have the proven quality track record.
So no, we really cannot ignore vulnerability reports.
The slop is gone
A sizeable portion of users seem to still believe we still suffer from a lot of made up and hallucinated reports by these AIs. We really don’t.
Almost all the stupidity and lies we saw in the vulnerability reports made by AIs in the past have ceased. Since spring 2026 the slop reports are practically gone. Our current challenge is much harder: high volume, high quality.
AI talks too much
They always use too many words. Every report is now hundreds of lines and the explanations are convoluted and detailed (and often repeat themselves multiple times).
In a good report, the human writes a first paragraph explaining the problem briefly and then they submit the full report. It’s annoying when the reporter surrender everything to the AI, including the communication part.
The AIs are also still bad at figuring out exactly where the line is between a bug and a vulnerability, and if it is a vulnerability, what severity it is. This is not a surprise though, as this is also the case for human reporters.
A good trend is that they often produce good reproducers or proofs of concept helping to trigger and prove the problem/vulnerability. That was a really complicated part of the reporters’ work back in the day, now it’s just a prompt or two.
Everything is with AI now
These days, like within the last few months, we have received hundreds of vulnerability reports submitted to curl and every single one had AI involved to some extent.
It is generally easy to tell on the language and the length and details of the report.
To be fair to the AI the AI written stuff is not nearly as repetitive and full of inaccuracies as the posts by people who as anti AI and have based their entire opinion on its capabilities on the early models from a few years ago. I can’t help but think of it as Anti-AI slop sometimes.
I read an interview with the developer of sqlite talking about how he had a bug apocalypse. He said he spent a year following a testing methodology in aerospace and that it got him out of that phase. Now he sells access to the test suite, which is a way to get payment for dealing with a ton of bugs. What I would add is that you can use AI to write aerospace quality testing too. It is just more code to read.
Hello Adrian,
please provide us the link to the interview to spare some CO2, data, etc pp and prevent others from using a search engine.
I was not able to find that interview. I just found the projects doc about their test suite. https://sqlite.org/testing.html
This isn’t an interview, and it is a video, but I completely separately found this, which includes a description of that history of bug avalanches followed by ways to stem the tide with testing. At the time this was recorded the tide of AI bugs was not yet stemmed.
https://youtu.be/V_qzqY1bb7I?is=pBFE9QV9vOepjU7j
The “high volume, high quality” part is the interesting one. A wrong report wastes an hour, a correct but 400-line report wastes an afternoon, and you can’t skip it. Have you considered asking reporters for a short human-written summary at the top, three sentences max, and sending back anything that doesn’t have one? Feels like the cheapest filter that still respects the people doing real work.
@Marya: the point remains the same. If the report identifies a flaw, or even a vulnerability, we want to fix that. It does not help us to dismiss the report because it was written badly by an AI. Dismissing does not remove the problem.