AI Bug Reports Swamp curl as Anthropic Launches OSS Scanner

Accurate reports, new problem

For years, the complaint from open-source maintainers about AI-generated security reports was that they were junk: confident, plausible, and wrong. Daniel Stenberg, who created and leads curl, the ubiquitous URL transfer library, now describes a different problem. Curl receives an AI-generated vulnerability report roughly every 18 hours, compared with about one report a week before AI tools became common. Most of them are now technically accurate. 3

Another account of Stenberg's comments puts the incoming rate at four to five times the 2024 level, double the 2025 level, and more than one report per day on average. 4 The figures differ slightly in framing but describe the same trend. Volume is climbing fast, and better quality has not reduced the workload.

The shift happened quickly. In early 2025, most AI reports reaching curl were hallucinated. By 2026 the models had improved enough that the reports described real issues. 3 That change shows up in the project's bug bounty decisions. Curl shut down its HackerOne program in February 2026, citing AI slop. It reopened the program a month later after quality improved, even though volume kept rising. 3

Finding is not fixing

Stenberg's central point is that AI has become very good at detection but remains weak at the parts of security work that take the most effort. According to his account, the models judge severity poorly and often propose incorrect fixes. That leaves triage, impact assessment, and remediation to human maintainers. 3 Duplicates add to the burden. Different researchers prompting the same models tend to get the same findings, so maintainers end up reviewing one bug several times. 3

Severity matters here too. Despite the flood, most of what is being found in curl is low or medium severity, and the project's last high-severity CVE dates to October 2023. 4 Many accurate reports describe real flaws that are minor. Each one still needs a human to verify it, rank it, and close it out.

Stenberg is also skeptical of the companies driving this wave. He has called Anthropic's Mythos model a "PR stunt." He argues that AI will not replace human jobs but could undermine bug bounties, because paying per report stops making sense when machines can produce valid findings cheaply and in bulk. 3

Anthropic's numbers tell the same story

Anthropic's own disclosure data points the same way. Its coordinated-disclosure dashboard shows Claude has surfaced 6,157 vulnerabilities across 591 open-source projects, but only 516 of those, about 8.4%, have been patched upstream. 5 Dark Reading argues that the real story behind the headline count of 129,000 vulnerabilities is the gap between discovery and remediation. It cites a SOCRadar warning that defenders could end up with a larger backlog rather than better security if finding keeps outpacing fixing. 5

Against that backdrop, Anthropic has launched OSS Scanner, an opt-in service that runs periodic security scans of enrolled open-source repositories using its strongest models, at no cost to the projects. 1 The company says the service grew out of its experience using Claude to find bugs during Project Glasswing. 1 Core maintainers sign up by opening a pull request that adds a project.yaml file to Anthropic's oss-scanner repository. The file includes a repo link, a primary contact, and a Dockerfile that builds the project so a fully offline agent can audit it. Anthropic says it manually verifies maintainer status before enrolling any project. 1

One review notes that the service targets established projects with critical infrastructure impact and that its scans include Claude Mythos. 2 It also stresses the main caveat. Reports are entirely model-generated and arrive without human review, so they may contain false positives, duplicates, and wrong severity ratings. Maintainers must validate every one before acting. 2 Anthropic paired the launch with a Critical Infrastructure Defense Program focused on operational technology security. 5

The reading

Taken together, these accounts point to a clear conclusion. AI vulnerability discovery has largely solved its accuracy problem, and the bottleneck has moved to the people who triage and fix bugs. OSS Scanner, by its design, hands that downstream work back to maintainers. Its opt-in structure and verification step are sensible guardrails. Even so, sending unreviewed reports to small volunteer teams risks adding exactly the kind of load Stenberg describes.

The tools that would help most are the ones the industry has not yet delivered well: reliable severity assessment, deduplication across researchers, and patches correct enough to merge. Until then, the 8.4% patch rate is the more useful number to watch. Better discovery alone does not make software safer if fixes cannot keep pace.