Anthropic OSS Scanner free vulnerability detection for open source

What changed

Anthropic has launched OSS Scanner, a free vulnerability detection service for critical open-source projects. The service runs the company’s strongest models, including Claude Mythos, against enrolled codebases on Anthropic’s infrastructure. The program began accepting projects on October 8, 2026.

Unlike traditional security audits, OSS Scanner reports bypass human review and go directly to project maintainers. This design choice accelerates the delivery of findings but transfers validation work onto development teams. In the first 11 days following launch, the scanner produced 77 critical and 803 high model-rated findings. By October 19, Anthropic had accepted 114 projects from roughly 650 applications.

The service operates on an opt-in basis. Core maintainers enroll by submitting a pull request to add their project configuration to the anthropics/oss-scanner GitHub repository. Required fields include a git repository link, primary contact email, and a Dockerfile that builds the project offline. Optional fields allow maintainers to specify a threat model file that defines what code should be tested and how severity should be classified, plus a GPG public key for receiving encrypted reports.

Anthropic conducts scans in isolated virtual machines with no internet connectivity. The Dockerfile builds with network access to install dependencies, but the actual security audit runs completely offline. When vulnerabilities are identified, reports include self-contained reproducers, explanations, root cause analysis, and candidate patches where available.

Why it matters

The absence of human review before reports reach maintainers creates distinct trade-offs. Anthropic’s penetration testers audited 97 critical and high findings across 48 projects. Of those, 85 findings (88%) met the company’s coordinated vulnerability disclosure standards. Only 1 was invalid. However, 11 reports duplicated known issues or other scan findings, suggesting maintainers may receive inflated severity counts.

Real-world results support the accuracy claim. WolfSSL, a cryptography library enrolled in the program, received 74 reports with 72 confirmed valid and 2 false positives. Five findings became CVEs. Curl, the file transfer utility, had a High-severity HTTP/2 upload use-after-free vulnerability identified by the scanner (CVE-2026-92392). The vulnerability affected versions 8.12.0 through 8.22.0 and was fixed in 8.23.0, released October 14, 2026.

For infrastructure-critical projects, free audits address a genuine bottleneck. Anthropic previously surfaced over 29,000 candidate vulnerabilities across six months. Human review triaged approximately 6,000 of those, while nearly 5,000 more were sent directly to maintainers who requested all findings unfiltered. OSS Scanner formalizes that second path for projects equipped to handle it.

Acceptance criteria mirror Google’s OSS-Fuzz program. Anthropic prioritizes established projects with critical infrastructure impact, exposure to remote attacks, and large user bases. The company evaluates OpenSSF criticality scores, update frequency, open-source license compliance, GitHub metrics, and security relevance to key sectors. Brand-new projects typically do not qualify, nor do forks unless they diverge significantly from the original.

Projects already overwhelmed with vulnerability reports may find the service counterproductive. Anthropic explicitly targets teams that already keep pace with verified high and critical reports and want deeper security coverage. Unlike traditional disclosure timelines, OSS Scanner imposes no 90-day deadline on unvalidated findings, shifting timeline pressure to maintainers.

What to test

Before enrolling, maintainers should verify several claims and consider operational impact.

• Accuracy baseline. Anthropic claims 88% of critical and high findings met its disclosure standards based on 97 audits. Compare OSS Scanner output against your own security knowledge to determine whether the scanner finds real issues in your codebase. Check whether findings duplicate your existing known issues or internal backlog.

• False positive rate. The audit showed only 1 invalid finding in 97, but real-world results will depend on your project’s specifics. Run a trial scan and measure how many reports require triage or dismissal. WolfSSL’s 2 false positives in 74 reports (97% precision) offers additional data, though your project may produce different results.

• Dockerfile reproducibility. Anthropic’s agents build your project in an offline environment. Test your Dockerfile locally to confirm it builds successfully and that the resulting environment works as expected. Any build failures will delay scans, and debugging occurs via email communication with Anthropic.

• Threat model clarity. If you provide an optional threat_model.md file, verify it communicates your actual security priorities. Without it, Anthropic’s models will guess at severity classification and in-scope code. Test whether the default output aligns with your security needs.

• Report velocity and triage burden. The first 11 days produced 880 total findings across 114 projects. Calculate whether your team can triage findings faster than they arrive. There is no enforced disclosure deadline, but reports go directly to your inbox unfiltered.

• Patch quality. Anthropic provides candidate patches where available. Test whether these patches compile, pass your test suite, and match your code style. Evaluate whether candidate patches reduce your workload or add review overhead.

The conclusion

OSS Scanner addresses a genuine infrastructure security gap. Free audits by strong language models can help critical open-source projects identify real vulnerabilities faster than manual review allows. The 88% accuracy rate on critical and high findings, combined with real-world CVE production, suggests the service delivers practical value.

The trade-off is substantial. No human review means maintainers absorb triage work and may see duplicates or inflated severity counts. Projects already struggling with vulnerability reports should avoid enrollment. Teams with capacity to validate and prioritize findings will realize more benefit.

Acceptance remains selective. Anthropic screens for infrastructure criticality and maturity, similar to OSS-Fuzz criteria. New projects, forks, and specialized libraries are unlikely to qualify. For projects that do get accepted, the cost-free audits from frontier models represent a meaningful security upgrade. Monitor whether the 88% accuracy holds as the scanner scales and whether maintainer feedback shifts acceptance standards or threat model practices.

If you maintain critical infrastructure code and have team capacity to handle fast-arriving reports, enrollment is worth testing. For teams already struggling with vulnerability management, consider improving your existing triage process first.

Related reading

• Anthropic sandbox-runtime: OS-level agent security
• OpenMobile-2 smartphone agents match proprietary rivals
• Claude Haiku 5.5 vs MiniMax M3.1 Flash