Anthropic Gives Open Source Maintainers a Fast Track to Mythos Vulnerability Findings

What it is

Anthropic's OSS Scanner is a free, opt-in service that runs Mythos and other Claude models over open source projects and sends maintainers reproducers and proposed patches directly. It skips Anthropic's human triage queue. Anthropic's human-reviewed disclosure program continues alongside it, mainly for projects that lack the resources to triage raw findings.

How it works

Eligible maintainers apply through the GitHub contribution process. The page does not give the repository or process name, only that route.

• Inputs from the maintainer: a project configuration and a Dockerfile that prepares the code for an offline audit. An optional threat_model.md sets attack surfaces, exclusions, severity criteria, and expectations for reports and patches. Its purpose is to stop the model flagging behavior the project explicitly permits.
• Controls: Anthropic verifies maintainer status. Projects can pause reports or withdraw.
• Disclosure clock: unvalidated reports do not start a 90-day clock. If Anthropic later validates a finding through its human-reviewed program, the 90 days start when the maintainer is notified. Anthropic may change the policy with advance notice and an opt-out.

Preconditions / Affected

The FAQ targets established projects with significant infrastructure or security impact. They should already keep up with verified high- and critical-severity reports and have capacity to investigate more. Projects that cannot are steered to the human-reviewed path.

What to do

If you maintain a qualifying project:

1. Write the offline-audit Dockerfile and project configuration.
2. Write a threat_model.md. It is optional, but it is the main lever against inflated severity and out-of-scope findings.
3. Plan remediation capacity before enrolling. Pause or withdraw if the volume is too much.

If you consume these projects, expect more fixes and releases. Users still have to install the fixed versions, a delay Stenberg calls out.

Caveats

• Volume vs. review: in six months Anthropic found more than 29,000 candidate vulnerabilities and triaged about 6,000. Nearly 5,000 unvalidated reports have already gone out at maintainers' request.
• Evaluation limits: pentesters checked 97 high/critical findings across 48 projects. 85 met disclosure criteria, 11 were real but duplicates, and 1 was invalid. This does not establish a false-positive rate for other severities or for projects that join later.
• Metric inflation: the true-positive metric on Anthropic's dashboard counts duplicates and issues maintainers may decline to fix, such as those outside the threat model or in unreachable code. The dashboard itself calls patch counts the more reliable but delayed measure.
• Maintainer pushback: maintainers have challenged inflated severity ratings and findings that misread their threat models.
• Reported outcomes:
• wolfSSL: 72 of 74 findings valid, 5 became CVEs.
• PostgreSQL: some patches were usable almost unchanged.
• Firefox: 271 fixes shipped in April, which took round-the-clock work.
• curl: AI tools had already driven roughly 200–300 fixes before its first Mythos report. Stenberg warns that a smaller later haul can mean earlier cleanup worked.
• Unevenness: the same output can produce a patch sprint at a well-resourced project and a backlog at a volunteer-run one. Anthropic admits it has not yet achieved a sufficient reduction in cyber risk, and often saw months between discovery and fix.