Anthropic has released comprehensive statistics from its coordinated vulnerability disclosure program, revealing how its models processed nearly thirty thousand vulnerability candidates before the public launch of the OSS Scanner service on 8 October 2026.
The dashboard snapshot taken on 2 October shows that models generated 29,439 candidate vulnerabilities. External security firms reviewed 6,123 of them (20.8 percent) and confirmed 5,674 as real issues, achieving a 92.7 percent confirmation rate among reviewed candidates. The remaining 23,316 candidates have not yet received independent verification.
Reports delivered to maintainers
Of the 6,157 reports sent to 591 open-source projects, 1,333 (21.6 percent) passed through external firm triage while 4,824 (78.4 percent) were delivered directly without the same level of human review. Maintainers responded to 5,103 reports, yet only 516 fixes (8.4 percent) have been merged upstream so far. Anthropic notes that the “fixed” counter reflects only changes known to the company and may undercount actual patches.
Why remediation lags behind discovery
- Patches take time; a September report may result in a November release that the 2 October snapshot does not capture.
- Many confirmed issues fall into “won’t fix” categories because they lie outside the project’s threat model or are unreachable in realistic scenarios.
- Even verified reports require maintainers to write tests, perform backports and evaluate behavioural changes before merging.
Claude Mythos reports now include self-contained reproducers, commit bisects and candidate patches when the model can generate them. The service runs inside isolated virtual machines with no internet access and is available only to projects that meet strict infrastructure and user-safety criteria.
Feedback published by Anthropic includes positive comments from maintainers of PostgreSQL, curl, OpenSSL, wolfSSL and HotCRP. On the academic CyberGym benchmark, the share of correctly identified vulnerabilities rose from under 20 percent to over 85 percent within a year. A small-scale review of 97 high- and critical-severity findings showed an 88 percent pass rate through the disclosure process.
The company acknowledges that some reports still overstate severity or misalign with individual project threat models. External firms agreed with Claude’s severity ratings exactly in 83 percent of cases and within one level in 97.1 percent of triaged reports.
