AI Bug Reports: curl's Reversal Exposes a Volume Crisis

From hallucinations to a firehose

For most of 2025, the complaint from open-source maintainers about AI-assisted security research was simple. The reports were fake. Language models invented vulnerabilities, fabricated proofs of concept, and wrapped it all in confident prose. In 2026 the complaint changed. The reports are increasingly real, and there are too many of them.

The curl project shows the shift most clearly. Its creator, Daniel Stenberg, says curl now receives an AI-generated vulnerability report about every 18 hours. Before AI, it was roughly one a week 1. Early AI submissions were mostly hallucinated, but by 2026 the models had improved enough that most reports were technically accurate 1. That sounds like good news, and partly it is. It also creates what Stenberg describes as a new kind of burden: every accurate report still needs a human to assess it, rank its severity, and write a fix 1.

What curl actually did with its bounty

Accounts of curl's bug bounty don't fully agree. One report says Stenberg shut down the project's HackerOne program in February 2026 because of AI slop. It says he reopened the program about a month later, once quality improved, even though volume kept rising 1. Another describes the paid bounty as ending permanently in January 2026 3.

The dates and the word "permanently" can't both be right. The broader direction is consistent, though. Infosecurity Magazine quotes OpenSSF CTO Christopher "CRob" Robinson, who points to people like Stenberg "doing 180 degrees" on AI tools. Robinson says maintainers now face a flood of genuinely relevant findings rather than junk 3.

The framing matters. Stenberg has publicly called Anthropic's Mythos a "PR stunt." He argues AI won't replace human jobs, but he thinks it could kill bug bounties 1. His reasoning is that AI is very good at finding bugs and poor at judging how serious they are or producing correct patches. That leaves the hardest work with the maintainers 1.

Duplicates: the hidden tax

A second problem is duplication. When many researchers point the same models at the same codebases, they find the same bugs. Stenberg notes that duplicates have become common because different people prompt the same AI and get the same answers 1.

Linus Torvalds made the same point about the Linux kernel. He said its security mailing list has become "almost entirely unmanageable," with heavy duplication from people using identical tools 2. His advice to would-be reporters: if AI helped you find a bug, someone else probably found it too 2.

This undercuts the economics of bug bounties. Bounties pay for being first to find something rare. If AI makes many findings easy and simultaneous, the reward stops matching the effort, and maintainers spend their time sorting identical submissions.

Slop hasn't disappeared

AI-generated junk is not gone, and here the picture diverges. GitHub's Jarom Brown says his team is inundated with submissions that show no real security impact. These include reports without proofs of concept, theoretical attacks that fall apart under scrutiny, and issues already on GitHub's published ineligible list 2.

Google's response went further. On October 1, 2026, it stopped accepting product vulnerability reports through its Open Source Software Vulnerability Reward Program after a wave of AI-generated submissions that were mostly wrong 5. The pause is narrower than the headlines suggest:

• Reports filed before the cutoff will still be processed.
• Supply chain reports are unaffected.
• Some Google Cloud-related repositories can still route bugs through the Cloud VRP.

Google has promised an update by the first quarter of 2027 5. Open Channels News described the situation as "security spam." Plausible reports now cost almost nothing to produce, while verifying each one remains expensive for the people receiving them 4.

The real bottleneck is human triage

These accounts can look contradictory. Curl says reports got better, while Google says they're mostly wrong. In my view, they describe the same underlying failure. Accurate or not, every AI-generated report goes into a queue that only a small number of people can work through. Whether the queue is filled with junk or with legitimate bugs, it outgrows the maintainers.

That is why industry coalitions are starting to organize around open-source security for the AI era. The problem is no longer finding bugs. It is staffing the work of confirming, ranking, and fixing them 3.

The curl story is less a story about AI getting smarter than a warning that discovery now scales and remediation does not. Bug bounties were built for a world where findings were scarce. Unless funding and tooling move toward triage and fixes, programs will probably keep pausing, restarting, and narrowing their scope as maintainers try to stay afloat.