Anthropic announced OSS Scanner on 8 October 2026 as a free, opt-in service offering periodic security scans to open-source projects. Each report includes an explanation, a proof of concept and a suggested fix where available, and Anthropic sends the model-generated reports without human review. Anthropic warns that some reports will contain inaccuracies, including incorrect severity ratings.
Anthropic says OSS Scanner is meant for projects with the capacity to keep up with the findings. Other projects continue to receive human-verified disclosures under Anthropic's coordinated vulnerability disclosure policy. The OSS Scanner documentation says the service is built for projects already able to keep up with verified high- and critical-severity vulnerability reports that want to further secure their code.
Anthropic says OSS Scanner is meant for projects with the capacity to keep up with the findings.
Enrolment, environment and report scope
Core maintainers enrol by opening a pull request that adds a project configuration identifying the repository, primary contact and environment-building Dockerfile. The Dockerfile builds with network access, but the subsequent security audit runs without internet access. The scanner team recommends checking that the project's tests pass inside the built container.
An optional threat-model file can describe code to test, adversarial inputs, exclusions, a severity rubric, report formatting, patch expectations and deduplication preferences. Without that file, the scanner still functions but makes its own guesses about corner cases, and maintainers can modify the file between scans.
Humans review sign-up requests and verify project security contacts. Automated checks confirm that reports concern the target project rather than its dependencies. Changes to a project's Dockerfile receive review before subsequent runs.
What arrives after the first scan
The initial scan produces a bundle of emailed bug reports, with agents double-checking bugs, proposing patches and performing root-cause analysis. Later scans look for newly introduced vulnerabilities and ones previously missed. The frequency of later scans may depend on pipeline size, project usage and other factors.
Anthropic will not impose a 90-day coordinated disclosure period on these unvalidated findings because they may contain false positives. If Anthropic later validates a report manually through its existing disclosure programme, it may disclose the finding under that policy starting 90 days after notifying the maintainer that a human has validated it.
Maintainers can temporarily pause reports through a configuration change. Opting out returns a project to receiving reports only through the standard coordinated disclosure process. Maintainers can reply to emailed reports with feedback about findings.
Validation remains the bottleneck
Anthropic's customer account says Comcast used Claude Mythos Preview to identify an authentication vulnerability in a public-facing platform during an assessment covering 258 business-critical systems and approximately 170 million lines of code. Comcast engineers validated the finding against the running application, confirmed the issue was reachable, traced the root cause, remediated it and checked for prior exploitation. Noopur Davis of Comcast is quoted saying, "Validation of these volumes of findings is the new bottleneck."
Davis distinguishes a finding from a validated vulnerability, and a vulnerability from one exploitable in its target environment. Booz Allen's Brad Medairy says teams still need to validate findings, deduplicate patterns and route issues to owners.
Anthropic says verification, prioritisation and fixing remain difficult, and that it often saw months pass between finding a vulnerability and fixing it in Project Glasswing. Its open-source plans include automating triage and patching for projects that want those services, while retaining human-verified disclosure for projects that need it.
Under Project Glasswing, Anthropic scanned hundreds of widely used open-source projects, had humans triage and review many potential vulnerabilities, and privately reported findings to maintainers. Some maintainers able to triage at scale asked for everything the models had found, reviewed or not, and Anthropic describes OSS Scanner as its response to those requests.
Anthropic's Cyber Verification Program announcement describes an expanded programme that makes advanced cybersecurity capabilities available to verified users. The announcement says generally available models can continue code review, patching known issues, finding vulnerabilities in owned source code and triaging security alerts. It also describes a Red Team Access tier for authorised penetration testing, limited to systems the organisation is authorised to test.
Uber CISO Philip Martin describes collecting real-time telemetry and taking real-time action within deployed agent software. He prioritises checking the tool an agent is about to call and the destination it is about to contact.
Related directory entries: