Anthropic OSS Scanner Sends Unreviewed AI Findings to Cut Vulnerability Patch Delays

Anthropic has launched OSS Scanner, a free opt-in service that uses its strongest Claude models, including Mythos, to identify vulnerabilities in open source projects. Participating maintainers receive reports containing reproduction steps and proposed fixes immediately, bypassing the company’s internal research review process.

What happened

Anthropic’s researchers previously struggled to validate the volume of findings generated by their AI models. Over a six-month period, the company identified more than 29,000 candidate vulnerabilities but manually reviewed and triaged only approximately 6,000 of them. Some maintainers requested access to the remaining unvalidated findings for their specific projects.

In response to these requests, Anthropic has already sent nearly 5,000 unreviewed reports. The OSS Scanner allows more open source projects to sign up for these findings without waiting for human validation. The company acknowledges that Project Glasswing uncovered vulnerabilities but failed to achieve a sufficient reduction in cyber risk due to long remediation times.

Why it matters

The delay between vulnerability discovery and patching has drawn criticism from security researchers. Anthropic states that months often passed before fixes were applied, creating a significant bottleneck. By providing unreviewed findings, the company aims to help projects turn AI-generated insights into actual code fixes more rapidly.

Anthropic tested an early version by having penetration testers examine 97 high- and critical-severity findings across 48 projects. Of those, 85 met coordinated disclosure criteria, while eleven were real issues that duplicated known problems. One finding was invalid. The company notes that maintainers have challenged inflated severity ratings and findings based on misunderstood threat models.

Next steps

• Evaluate your team’s capacity to handle a high volume of unvalidated vulnerability reports.
• Review proposed fixes for accuracy before integrating them into your codebase.
• Monitor for false positives that may arise from AI misinterpreting your specific threat model.
• Prioritise findings that meet your project’s coordinated disclosure criteria.

Detection notes

Security operations teams should monitor for unusual spikes in vulnerability reports from automated sources. Check whether incoming alerts include detailed reproduction steps or just generic warnings. Look for patterns where multiple findings target the same code module, which may indicate a systematic AI error. Verify the severity ratings against your internal threat model before escalating. Contain potential risks by isolating affected components during manual review. Ensure that any proposed patches are tested in staging environments before deployment to production systems.

What to do and how to stay safe: Anthropic

• Audit your open source dependencies to identify projects using AI-driven vulnerability scanning.
• Establish a process for manually validating AI-generated findings before applying patches.
• Monitor vendor dashboards for updates on the accuracy and false-positive rates of these tools.
• Once the vendor provides an update, review any changes to the scanning methodology or reporting format.

Step-by-step guide: Emergency Patching: How to Fix Critical Flaws Without Breaking Production

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

Frequently asked questions

How many vulnerabilities did Anthropic find before launching the scanner?

Anthropic found more than 29,000 candidate vulnerabilities over six months but manually reviewed only 6,000.

What was the accuracy rate of the early scanner test?

Of 97 high- and critical-severity findings examined, 85 met disclosure criteria, eleven were duplicates, and one was invalid.

Which projects have used these findings?

Mozilla shipped fixes for 271 vulnerabilities, and wolfSSL reported that 72 of 74 findings were valid.