OSS Scanner

• Enrolling a widely used parsing or networking library that handles untrusted input to get periodic deep scans by frontier models
• Getting raw vulnerability reports with reproducers and candidate patches as soon as they are generated, instead of waiting for Anthropic's human-reviewed disclosure queue
• Writing a threat model and severity rubric so the scanner focuses on in-scope code and rates issues the way the project does
• Feeding reports with reproducers into an existing security triage and patch process to fix issues before they reach a release
• Crediting fixes to the scanner by citing the report ID (for example ANT-2026-ABCD1234) in commit messages, which Anthropic requests but does not require

OSS Scanner turns frontier-model vulnerability research into something an open-source project can simply sign up for, which matters because the maintainers of critical code often lack the budget for tools like this. The design is honest about its tradeoff: you get reports quickly because no human has checked them, so the value depends on whether your team can triage a stream of plausible but sometimes wrong findings. Investing in a good threat model and severity rubric is the most useful thing a maintainer can do before enrolling. Projects that are already stretched are better off staying with Anthropic's human-reviewed disclosures.

About OSS Scanner

OSS Scanner is a hosted service, not a tool you install. Anthropic's Frontier Red Team runs the pipeline on its own infrastructure, and maintainers interact with it through a pull request and email. It launched on October 8, 2026, as part of the Anthropic Cyber Mission, and is inspired by Google's OSS-Fuzz, whose eligibility criteria it borrows. To enroll, a core maintainer opens a pull request to the public anthropics/oss-scanner GitHub repository that adds a projects/<name>/project.yaml (the git repository to clone, which does not have to be on GitHub, plus a primary contact email), a Dockerfile that installs every dependency and builds the project, and optionally a threat_model.md. Anthropic manually confirms that the person is a core maintainer and accepts projects case by case, using OSS-Fuzz-style criteria: critical impact on infrastructure and user security, exposure to remote attacks (for example, libraries that process untrusted input), and the number of users and dependent projects. After the merge, the scanner builds the project with network access in an isolated VM, then cuts off Internet access and scans it. Agents in the pipeline double-check bugs, analyze root causes and propose patches. Each report includes a self-contained reproducer, an explanation of the vulnerability (with a bisection to find when the bug was introduced, where possible) and a candidate patch when one is available. The first scan is followed by periodic rescans, whose frequency may depend on how many projects are enrolled, how widely a project is used and other factors. The threat model file is the main control maintainers have: it can define what is in scope, which inputs are adversarial, a severity rubric, report and patch style, and deduplication granularity. The key tradeoff is speed against verification. Anthropic already sends human-reviewed reports through its coordinated vulnerability disclosure (CVD) process, but that review is slow. OSS Scanner is the fast track: reports are fully model-generated and sent without human triage, so some will be wrong, for example with an inflated severity rating or a misread threat model. In Anthropic's own validation of an early version, its penetration testers found that 85 of 97 critical and high-severity findings, across 48 projects, met the bar for its CVD process. Because findings are unvalidated, Anthropic places no 90-day disclosure deadline on them; if it later validates a report manually through its CVD program, it may disclose that vulnerability 90 days after notifying the maintainer. It is free, and Anthropic says it covers the full cost. It suits teams that can triage at volume, not projects already overwhelmed by reports.