Anthropic bypasses human review to send raw AI vulnerability reports directly to OSS maintainers

Anthropic has launched OSS Scanner, a free, opt-in service that delivers unreviewed vulnerability findings from its AI models directly to open source maintainers. The tool uses Claude models, including Mythos, to scan code and generate reports with reproduction steps and proposed fixes, bypassing the company’s internal manual triage process.

The shift aims to address a bottleneck in remediation timelines. According to Socket, Anthropic discovered more than 29,000 candidate vulnerabilities over six months but manually validated only approximately 6,000. Some maintainers requested immediate access to remaining findings, prompting Anthropic to send nearly 5,000 unvetted reports to those projects.

Context and Background

Anthropic stated in its Cyber Mission announcement that Project Glasswing had not yet achieved a sufficient reduction in cyber risk. The company acknowledged that months often passed between vulnerability discovery and patching, a delay that drew criticism from security researchers. OSS Scanner is designed to help projects turn AI findings into fixes faster by removing the wait for researcher validation.

The service targets projects capable of handling high-volume reports. The extended FAQ notes that the tool is best suited for teams that can already manage verified high- and critical-severity findings and have the capacity to investigate additional, unvalidated issues. This approach marks the first step in a broader effort to automate triage and patch development.

Who is Affected

Early evaluations suggest mixed results regarding accuracy and severity. Penetration testers examined 97 high- and critical-severity findings across 48 projects; 85 met Anthropic’s disclosure criteria, while one was invalid and eleven were duplicates. Anthropic acknowledges that some maintainers have challenged inflated severity ratings and findings based on misunderstood threat models.

Specific projects have reported positive outcomes from early access. PostgreSQL maintainer Noah Misch said some proposed patches were usable with minimal changes, helping resolve issues before general availability. wolfSSL’s Todd Ouska reported that 72 of 74 findings were valid, with five resulting in CVE assignments.

Mozilla’s Firefox team shipped fixes for 271 vulnerabilities identified during an early Mythos Preview evaluation in April. Mozilla engineer Bobby Holley noted that the work required sustained engineering effort and warned that other projects might need to reprioritize tasks to handle such volumes.

What should you do?

Open source maintainers considering this tool should assess their team’s capacity to triage unverified reports. Projects should verify that their threat models align with the findings before applying proposed fixes. Security teams should monitor for inflated severity ratings and validate all AI-generated claims against their specific codebases.

What to do and how to stay safe: Anthropic

• Assess your team’s capacity to investigate and triage a high volume of unverified vulnerability reports before opting in.
• Validate all AI-generated findings against your specific threat model and codebase context before applying proposed patches.
• Monitor for inflated severity ratings or false positives that may not align with your project’s actual risk profile.
• Coordinate with other maintainers to share insights on the accuracy and usability of AI-proposed fixes for your specific stack.

Step-by-step guide: Risk-Based Vulnerability Management: Prioritize Fixes by Actual Danger

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Does Anthropic review the vulnerability reports before sending them to maintainers?

No, the OSS Scanner sends reports without waiting for Anthropic’s researchers to validate the findings, allowing maintainers to act faster.

How accurate are the AI-generated vulnerability findings?

In early testing, 85 of 97 high- and critical-severity findings met disclosure criteria, though some maintainers reported inflated severity ratings.

Which open source projects have used this tool?

Projects including PostgreSQL, wolfSSL, Mozilla Firefox, and curl have received findings or participated in early evaluations of the technology.