Lab program data, vendor accounts of review queues and maintainers’ reports, not a systematic field-wide patch-rate measurement. Microsoft distinguishes internal validation from maintainer acknowledgement or acceptance; Anthropic’s counts use partial partner reporting with heterogeneous triage and incomplete patch data. The Cyber Mission describes delivery and remediation efforts, not a measured reduction in risk.
Corroborated: Supported by at least two independent sources.
Evidence
Complaints centre on duplicate and low-quality AI-assisted reports straining triage.
75 of about 530 disclosed high/critical open-source bugs were patched at the time of the update; Anthropic notes this was early in its 90-day window and likely undercounted.
Microsoft identifies validation and review queues as constraints, and distinguishes 155 internally validated open-source reports from 93 with maintainer acknowledgement or acceptance; neither is a count of shipped patches.
Anthropic says months often passed between finding and fixing Glasswing vulnerabilities and describes opt-in unreviewed scanning for maintainers able to handle the volume.
Anthropic’s expanded program counts rely on partial partner reporting, heterogeneous triage and incomplete patch data; they do not establish a comparable patch rate.
The pipeline separates candidate findings, dynamic reproduction, validated exploits and candidate patches; generating a candidate patch is not evidence of upstream delivery.
Adds remediation evidence: Trail of Bits reports delivered stable-kernel backports for 149 distinct CVEs, with human review and maintainer acceptance still required.
Qualifies generalization: an observational preprint finds no measured remediation slowdown among fixed-and-disclosed flaws, but cannot exclude an undisclosed backlog. The exploited-share decline is suggestive, and count estimates do not establish fewer exploited vulnerabilities.
Key questions that rely on this finding
- Is AI shifting the balance between finding and fixing vulnerabilities?Discovery is ahead. Verification, triage, coordinated disclosure, safe repair and getting fixes to users remain bottlenecks between AI findings and protection.