Anthropic launched OSS Scanner yesterday: free, periodic vulnerability scans for critical open-source projects, generated by their strongest models and sent to maintainers without human review in between. I am quoted in the announcement.
Eighteen months ago the AI-generated reports reaching the OpenSSL Security Team were appalling: high volume, little substance. Today, by our own numbers, 75 percent of external vulnerability reports to the OpenSSL Library show strong AI authorship. Over the last two years they were accepted as security issues at 22 percent against 20 percent for reports written by people, and rejected as not-a-bug less often, 26 percent against 38 percent. The reports we received from Anthropic, raw model output included, were as good and sometimes better than what we get from people.
The number of reports written by people has not left its historical range; the growth is AI. Verification is still done by people. Every report, whatever wrote it, is triaged by the OpenSSL Security Team as a security issue, a bug, or not a bug before anything is fixed. When it comes with a working exploit, that review is immediate. We had to grow the team to keep up.
Anthropic says its models found 29,000 candidate vulnerabilities in six months and humans could review 6,000. The constraint now is the number of findings nobody has reviewed and where they end up. Findings that reach the maintainer get fixed once, upstream, for everyone. Findings that stay with whoever found them get patched in private forks. The OpenSSL Library will enroll in OSS Scanner for that reason.