Anthropic's AI Bug Hunt Has 37,966 Findings, but Its Public Ledger Lists Only 217 Fixes So Far

Anthropic's AI can produce bug reports faster than humans can check them. Seven months into Project Glasswing, VulnCheck's November 9 review counts 37,966 findings in Anthropic's latest accounting, but just 217 marked fixed. The gap puts the work of proving and repairing those bugs at the center of the story.

A finding is the start of the job

Researcher Patrick Garrity separately tracks 442 CVEs credited to Anthropic, with two identified as exploited in the wild. Those numbers measure different things: a candidate finding is not necessarily a confirmed vulnerability, and a CVE is not a count of attacks.

The public disclosure ledger, checked on November 11, still carries a November 2 snapshot. It shows 217 fixed entries, 1,654 duplicates and 419 withdrawn entries. Those are recorded statuses in that snapshot, so the figures cannot establish how many additional fixes maintainers may have completed since.

Garrity's criticism is about the widening queue. His September-to-November comparison shows 11,813 additional findings against just 15 additional recorded fixes. Discovering a possible defect does not close it: someone still has to reproduce it, assess its impact and decide whether the proposed patch works.

Anthropic acknowledges the human bottleneck

Anthropic's October announcement of OSS Scanner already described that problem. The company said it had manually reviewed about 6,000 of more than 29,000 candidate vulnerabilities, and had sent nearly 5,000 unverified reports in bulk to maintainers who requested them.

The opt-in scanner sends model-generated reports without human triage. Anthropic warns that some can be invalid, while also publishing positive feedback from participating projects and saying it will continue human-verified disclosures.

Garrity points to a less flattering sample from curl: of 34 published reports he reviewed, 19 had been rejected as vulnerabilities, six confirmed as low severity and nine still awaited assessment. That sample cannot establish a false-positive rate for the whole program, especially while cases remain unresolved.

It does show why a bigger discovery count is an incomplete measure of progress. The people receiving the reports have to do the validation, and the public record still leaves a large distance between finding something and recording it as fixed.

Some offers on this page may be paid placements or contain affiliate links.