What does Anthropic's new OSS Scanner actually do for open-source maintainers?

The scanner runs on model output alone

OSS Scanner is a new vulnerability-finding service built for open-source maintainers. It uses Anthropic's strongest models, including Claude Mythos, to scan codebases and produce reports that are fully model-generated. No human reviewer touches the output before it reaches the project team. This design enables faster and more frequent scanning than a traditional coordinated disclosure process allows. The trade-off is explicit: reports may be incorrect or invalid. Anthropic frames this as a fast-track option for maintainers who want raw findings immediately rather than waiting for human triage.

The service draws on experience from Project Glasswing, where Anthropic used Claude to hunt vulnerabilities in critical software. Over the last six months that effort produced over 29,000 candidate vulnerabilities. Human reviewers managed to validate approximately 6,000 of them. The gap between discovery and verification became the bottleneck. Meanwhile nearly 5,000 unverified reports were sent directly to maintainers who asked for everything the models had found. Exploits can now be developed in minutes, so the argument for speed is straightforward: projects that find and fix vulnerabilities faster can defend against attackers racing to weaponize the same weaknesses.

Benchmarks and early validation tell a specific story

On CyberGym, an academic vulnerability-finding benchmark, large language models have gone from finding under 20 percent of vulnerabilities at the beginning of last year to finding over 85 percent this year. That trajectory changed the quality of incoming bug reports for maintainers from mostly slop to high-quality submissions. Anthropic tested an early version of OSS Scanner with dozens of open-source projects. Initial disclosures contained hundreds of bug reports. Multiple vulnerabilities were chained into unauthenticated remote code execution exploits affecting those projects.

Each report ships with a self-contained reproducer, an explanation that includes a bisection to determine when the bug was introduced where possible, and a candidate patch when available. To validate the pipeline Anthropic asked the expert penetration testers who review its coordinated vulnerability disclosure findings to check 97 critical and high-severity vulnerabilities from the scanner across 48 projects. Of those 85, or 88 percent, met the bar for the CVD process. Eleven were real but duplicated known issues or other findings from the same scan. One was invalid, a false positive. Maintainers have seldom told Anthropic a high or critical finding was invalid. Some have said severity ratings can be inflated or that the scanner misunderstood the project's threat model.

Maintainers describe the signal as high

Feedback from early participants is concrete. Noah Misch of PostgreSQL said an unusually high fraction of findings uncovered defects, several reports came with fixes usable nearly as-is, and fast-track access let the team address issues before a general availability release. Anton Arapov of OpenSSL Corporation said early AI reports 18 months ago were appalling, but the reports received from Anthropic, raw model output included, were as good and sometimes better than what they get from people, particularly when a real exploit is attached. Todd Ouska of wolfSSL said of 74 reports all but two were valid, five became CVEs, and with patches attached the reports slotted right into their existing verification process. Eddie Kohler of HotCRP called the reports thorough and clear with strong understanding of the project's complex permission model and good bug prioritization. Daniel Stenberg of curl said OSS Scanner helped find multiple issues worthy of addressing, including one of the worst curl vulnerabilities reported in the last few years.

Enrollment follows an OSS-Fuzz-like model

Core maintainers of eligible projects enroll by submitting a pull request to a GitHub repository following a standard project template with additional guidance in an extended FAQ. Eligibility criteria mirror what OSS-Fuzz uses: projects should have a critical impact on infrastructure and user security, and decisions are made case by case. Anthropic says it will keep refining the system based on maintainer feedback and as models improve. It cannot guarantee the scanner will be perfect. For projects without the resourcing to triage reports themselves Anthropic will continue manually disclosing human-verified vulnerability reports via its existing CVD process.

Two related programs sit alongside the scanner. The Cyber Verification Program makes advanced cyber capabilities and reduced blocking classifiers available to qualifying security professionals. Claude for OSS provides free Claude Max 20x subscriptions to help remediate vulnerabilities and improve open-source projects. Neither program replaces the scanner; they address different parts of the workflow. The scanner finds issues. The verification program helps experts validate them. The subscription program gives maintainers compute to fix them.

What the model cannot promise

The scanner does not replace human review. It produces candidate patches, not guaranteed fixes. Severity ratings may be inflated. The threat model inferred by the model may not match the project's actual threat model. False positives exist, though at a low rate in the validation sample. The service is opt-in precisely because the output is raw. Maintainers who cannot triage high-volume unverified reports should stay with the CVD channel. The document does not say how many projects have enrolled so far, nor does it give a timeline for expanding eligibility beyond the current case-by-case process. It also does not disclose the false-positive rate for medium or low-severity findings, only the single invalid finding among 97 critical and high cases.