Counts are mostly reported by the organizations that built the systems; individual CVEs confirm specific cases. Google Threat Intelligence Group's September 2026 profile of likely AI-discovered vulnerabilities rests on its own heuristic for attributing discovery to AI.
Corroborated: Supported by at least two independent sources.
Evidence
GTIG profiles the vulnerabilities it judges likely AI-discovered: they skew toward Medium severity (58% against 28%) and remote code execution (50% against 26%) on GTIG's own risk ratings.
Microsoft reports Windows CVEs and internally validated open-source reports, separating those from documented maintainer acknowledgement or acceptance.
Anthropic reports vulnerability counts from partial Glasswing partner surveys and its own open-source work; heterogeneous triage and incomplete patch reporting limit comparison.
Calif found one of two critical PHP flaws (CVE-2026-109618 and CVE-2026-109619) using GPT-5.6 Sol, per OpenAI, and Wiz independently found both using frontier models; the PHP Foundation says one traces to a 2024 audit whose fix stalled.
Anthropic describes selected chains in open and proprietary software, including patched Chrome and Linux flaws; the demonstrations use an unreleased model and research assistance.
Google Project Zero reports that BigSleep produced 45 actionable V8 vulnerability reports from April to September 2026, against 37 for the Fuzzilli fuzzer, with about 25% overlap.
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.
- Can AI agents defend and oversee systems on their own?Not reliably. Some designs improve benchmark triage, but autonomous defense lacks operational validation and production oversight of agents has documented gaps.