Anthropic says its models identified more than 29,000 candidate vulnerabilities over six months, but researchers manually reviewed only about 6,000, making human validation a bottleneck. Each report includes a self-contained reproducer, an explanation of the suspected flaw, the change that introduced it where possible, and a candidate patch.
In an early test, penetration testers assessed 97 critical and high-severity findings across 48 projects, and 85 met Anthropic's coordinated disclosure criteria. Anthropic acknowledges that some maintainers saw inflated severity ratings. wolfSSL reported that all but two of 74 reports were valid and five became CVEs. Daniel Stenberg of curl attributed a high-severity HTTP/2 use-after-free, fixed in curl 8.23.0 as CVE-2026-92392, to an Anthropic researcher using the scanner.
Enrolment happens through a GitHub pull request containing project configuration and a Dockerfile, and scanning runs without internet access after setup. Anthropic checks applicants manually. Unvalidated findings carry no automatic 90-day disclosure deadline, while manually validated ones can enter coordinated disclosure. Human-reviewed disclosures remain available.
Why it matters if you build agents
Maintainers of libraries that agents depend on may start receiving unreviewed AI-generated reports, so triage capacity and validation will matter. Teams building agents on open-source components should expect more upstream fixes and plan how quickly they can adopt them. Check severity ratings independently before acting on them, since Anthropic has acknowledged inflated ratings.
Get stories like this every weekday
One short email with the day's agent tooling news.