The report
Anthropic reports that a Claude-assisted reachability review across 25 software projects found 493 issues suitable for potential attack-surface reduction. Its gated subset contained 11 applications, and researchers counted 93 unnecessary pre-authentication surfaces, 37 places where code ran or state changed before authentication, and 61 configuration combinations that weakened an intended gate. These buckets are findings in a research exercise, not confirmed intrusions or necessarily independently exploitable CVEs.
The study measures a different defensive goal from asking a model to generate more vulnerability reports. If software no longer exposes unused endpoints or unnecessary pre-authentication code, vulnerabilities in that unreachable code become harder to exercise. Reachability reductions still need regression tests and feature ownership, and a vendor-run evaluation on chosen projects does not prove that every suggested removal was safe to deploy. Anthropic research.
Three different measurements, not 493 confirmed exploits
Anthropic further distinguishes 162 configuration combinations whose interactions were security-sensitive: 94 silently disabled a security control and 68 produced inconsistent behavior within the same mechanism. The separate gated-project findings include 93 pre-authentication surfaces, 37 pre-authentication execution or state changes and 61 combinations weakening a configured gate. These are overlapping investigative categories in the authors’ sampled projects and cannot simply be added to a count of live exploits. The critical test after narrowing reachability is whether a formerly accessible code path is actually inaccessible in the supported configurations, without silently breaking legitimate uses. Primary research.
What those 493 findings measure
The 493 issues span the full 25-project exercise, whereas the 93 pre-authentication surfaces, 37 early execution/state effects and 61 weakened gate combinations refer to narrower categories discussed by the authors. Adding these figures produces neither a verified vulnerability count nor a measured reduction in compromises. The defensible result is an analysis of paths and configurations that may be unnecessarily exposed. Whether a proposed gate actually becomes mandatory needs deployment-specific testing; reducing a reachable path differs from proving the underlying code correct. Anthropic's methods and counts; review-capacity comparison.