Daniel Stenberg wanted people to tell him when they found security problems in curl, a small piece of software that helps computers transfer data across the internet. Since April 2019 the project had paid researchers for useful discoveries through a program on HackerOne, a bug-reporting platform. Then the reports became their own problem.
On January 26, 2026, Stenberg announced that the payments would stop at the end of the month. He named three trends that together forced the step: AI-generated "slop" reports, lower-quality human reports, and reporters who seem keener to find flaws than to help fix them. Before 2025, he wrote, about 15 percent or more of submissions were confirmed vulnerabilities. From 2025 that share fell below 5 percent. Curl still wants security reports. Reporters are asked to use GitHub's private vulnerability reporting or email, no money changes hands, and people who send slop will be banned. 1
That decision illustrates a problem that reaches well beyond one project. AI can help more people inspect software, and that can be good news for security. But every finding has to go somewhere. A person must decide whether the problem exists, whether an attacker can use it, and what to do about it. The easier it becomes to send a convincing report, the more that work grows.
A helpful discovery comes with a bill
A vulnerability report is not simply a message saying that something looks suspicious. A useful one tells a maintainer where to look, what conditions make the problem appear, and how someone can reproduce it. A maintainer can then investigate the risk and develop a fix.
An AI system can produce a polished explanation even when the underlying claim is wrong. It can also find a genuine bug while exaggerating what an attacker could do with it. Time spent chasing the wrong problem is time a small team cannot spend on another one. The CNCF, the foundation that hosts many open-source cloud projects, says low-quality reports can take hours to days to evaluate. 2
Stenberg did not blame every bad report on AI, and curl's experience was not typical. He wrote that over the previous four quarters curl's report volume on HackerOne rose sharply while comparable programs, including Ruby, Node and Rails, stayed flat or declined slightly. The bounty had also helped curl: it produced 87 confirmed vulnerabilities and more than $100,000 in rewards. Ending the payments meant giving up an incentive that had delivered real benefits as well as noise. 1
The rules are being rewritten
The Linux kernel, the core software under Android phones, servers and countless other systems, now gives reporters explicit rules for AI-assisted findings. Many AI-generated reports, it says, "lack an understanding of the kernel's threat model", the boundaries the system is meant to enforce, and invent theoretical consequences. It asks reporters to stick to verifiable facts, to supply a tested reproducer, to keep reports concise and in plain text, and to propose a fix. It also says that a bug found with AI assistance must be treated as public, so reporters should not post a reproducer openly and should offer it to maintainers on request. 3
The threat-model point is easy to miss. A program doing something surprising is not automatically a security breach. The question is whether it lets someone cross a boundary they should not be able to cross.
The CNCF gives the same advice from both sides. Maintainers should write their threat model down and commit it to the project, so reporters can check findings against it, and should triage in two passes: a quick check against the threat model, then full reproduction. Bug finders should write and test a proof of concept that uses the project's real interfaces and have a human review it before filing. Its examples of poor reports include proofs of concept that are just unit tests, ones that do not compile, and duplicates from the same reporter. Its rule is blunt: if you cannot follow these principles, do not file reports. 2
More findings do not automatically mean more safety
The next version of the problem may not be mostly false alarms. Better tools can surface real defects, too. Then the bottleneck moves: fixes must be reviewed, released, and installed by people who depend on the software. A discovery that never reaches a working update is only part of a solution.
Google's threat-intelligence group counted vulnerabilities disclosed per month rising from 5,045 in January 2026 to 10,477 in July and 10,740 in August. It cautions against reading that as pure AI discovery: automated assignment of identifiers can inflate the totals, and descriptions mentioning the Linux kernel alone generated about 5,000 CVEs over those eight months, none of them an exploited zero-day. Only 0.23 percent of disclosed vulnerabilities in 2026, about 1 in 431, had been seen exploited. The group calls the link between AI discovery and the increase an early indicator rather than an established trend, though it found that half of AI-discovered flaws allowed remote code execution, against 26 percent across all CVEs. 4
There is no reason that security research should be reserved for people who learned it the old way. A newcomer with an AI assistant might find something an expert missed. But the old obligation to the person on the receiving end still matters: check what you can, explain the evidence, and treat their time as a limited resource.
The promise of AI-assisted bug hunting is that more people can help secure the software everyone depends on. The test is whether the help arrives in a form the people doing the fixing can use.
References
References
- Stenberg, D. (2026, January 26). The end of the curl bug-bounty. https://daniel.haxx.se/blog/2026/01/26/the-end-of-the-curl-bug-bounty/
- Castle, G. (2026, April 16). The AI-driven shift in vulnerability discovery. https://www.cncf.io/blog/2026/04/16/the-ai-driven-shift-in-vulnerability-discovery-what-maintainers-and-bug-finders-need-to-know/
- Linux kernel security bugs: responsible use of AI. (n.d.). https://kernel.org/doc/html/next/process/security-bugs.html
- Vulnerability Discovery and Exploitation Trends in the AI Era. (n.d.). https://cloud.google.com/blog/topics/threat-intelligence/vulnerability-discovery-and-exploitation-trends-in-the-ai-era