Anthropic's models are finding open-source bugs much faster than those bugs are getting patched. A report can reach a maintainer and still sit there, with no fix you can pick up.
Anthropic's disclosure dashboard, current through October 2, covers Cyber Mission scans from November 1, 2025. A merged upstream patch is a change in the project. You still have to check the copy in your build, your image, and the lockfile you shipped last quarter.
Where the reports stall
| Stage | Count | Note |
|---|---|---|
| Discovered | 29,439 | Model findings through October 2 |
| Reviewed by outside firms | 6,123 | 5,674 confirmed valid |
| Reported to maintainers | 6,157 | 591 projects, most sent direct |
| Acknowledged | 5,103 | A maintainer replied |
| Patched upstream | 516 | A fix landed in the project |
The reviewed row and the reported row are different cuts of the pile. Outside firms had checked 6,123 findings. Maintainers had been sent 6,157. Of those reports, 1,333 went through triage and 4,824 went out direct. Direct reports may contain false positives. The reviewed slice is the clean part. On the findings outside firms actually checked, the true-positive rate was about 92.7%.
That confirmation rate is the encouraging figure. The patch rate is what changes the software you ship. The share of reported findings with an upstream fix rose from about 6.1% in May to about 8.4% by the October cutoff. The rate improved. A typical report still has no upstream fix. You can acknowledge a bug, file it, argue about severity, and ship nothing. A reply is much more common than a patch, and it means a person answered. A release is what you can deploy, and a version bump upstream still has to reach the machines you run.
If you maintain a library, read the funnel as a backlog. Tens of thousands of model findings were in the pile by October 2. A few thousand were checked by an outside firm, and nearly all of those held up. A similar number were handed to projects, often without that check. 5,103 of those handoffs drew a human response. 516 became a fix in the project. Leads are arriving faster than patches.
The discovered count will keep showing up in headlines. If you ship the dependency, track the patched count on the projects you can't replace. Between a lead and a fix, someone has to reproduce the bug, decide it matters in that project's threat model, and land a change other people will upgrade to. A suggested fix in the report doesn't finish that work. A high severity rating, or a threat model the scanner got wrong, can still cost you an afternoon when the underlying bug is real. Waiting for the patched count to catch the discovered count means waiting through releases that don't include these fixes.
Reports that skip a person
OSS Scanner is the opt-in lane for projects that want the unreviewed stream. Enrolled repos get periodic scans from Anthropic's strongest models, at no cost. Each report is supposed to include a way to reproduce the bug, an explanation, and a suggested fix when the model has one. The model writes that text, and it goes out with nobody checking it first, so your cost is the time to review it. The expected true-positive rate from here is above 90%, and the suggested fixes are still being improved. Projects that can't absorb that pace stay on the human-reviewed path.
Pentesters reviewed 97 critical and high findings across 48 projects. Of those, 85 met the bar for coordinated disclosure, 11 were real bugs that duplicated a known issue or another finding from the same scan, and one was a false positive. Maintainers have also seen severity ratings come in high, and they have seen the scanner misunderstand a project's threat model. wolfSSL maintainer Todd Ouska said that of the 74 reports his project received, "all but two were valid, and five became CVEs."
A valid report is still a lead. Those five CVEs are five patches someone has to land, ship, and get into the builds that depend on wolfSSL. A true-positive rate above 90% tells you the model is mostly pointing at real bugs. Whether those bugs are fixed on the machines you run is the question the patched count is there to answer.
If you depend on the code and you don't maintain it, those reports won't land in your inbox. Watch whether the projects you can't afford to have wrong show up as enrolled. When a model-written report surfaces, in a commit, a mailing list, or a CVE thread, treat it as a lead until a person confirms it. File a CVE, or ship a hotfix, only after someone reproduces the bug. When an upstream fix does land, check the version you actually run. The patched tally and the install on your machines are different problems, and only the install changes your risk.