Scope - What This Guide Covers
This guide helps CISOs understand and respond to data about AI-discovered vulnerabilities. It compares exploitation patterns of AI-discovered vulnerabilities with traditional methods, examines the timeline for known exploited vulnerabilities (KEVs), and offers practical response strategies for your security team. Use this guide when considering AI security tool investments, setting patch management timelines, or briefing your board on AI-related cyber risks.
Key Concepts and Definitions
AI-Discovered Vulnerability: A software vulnerability found using machine learning models or AI-assisted tools, rather than manual code review or fuzzing. Anthropic's Project Glasswing is an example.
Known Exploited Vulnerability (KEV): A CVE with confirmed active exploitation in production environments. The CISA KEV Catalog is the authoritative source for federal agencies, while commercial catalogs like VulnCheck track broader exploitation evidence.
Time-to-Exploitation: The time from CVE publication to the first observed exploitation attempt. This metric is now more critical for planning than theoretical CVSS scores.
Zero-Day KEV: A vulnerability exploited on or before its CVE publication date, indicating attackers discovered it independently or had advance knowledge.
The Exploitation Rate Reality
Data contradicts the "vulnpocalypse" narrative. Of 1,061 vulnerabilities attributed to AI-assisted discovery, 14 have been confirmed as exploited in the wild, a 1.3% rate, matching the overall exploitation rate for all vulnerabilities during the same period.
Anthropic's Project Glasswing reported over 23,000 findings, yet only 126 resulted in published CVEs, and only one has been exploited. This suggests AI tools generate many theoretical vulnerabilities, most of which can't be weaponized or aren't targeted by attackers.
Your takeaway: AI isn't creating a new class of highly-exploitable vulnerabilities. It's identifying more potential weaknesses, but exploitation still depends on attacker economics, access patterns, and target value calculations.
Requirements Breakdown: What's Changed
Exploitation Timeline Compression
The median time from CVE publication to KEV status dropped from 120 days in 2025 to 80 days in the first half of 2026. That's a 33% reduction in your response window.
However, zero-day and one-day KEVs decreased slightly: 23.43% of 2026 KEVs showed exploitation on or before publication day, down from 28.93% in 2025. Early exploitation (within 31 days) remained steady at roughly 200 CVEs.
Translation: You have less time overall, but the proportion of instant-exploitation events hasn't increased. Your patch prioritization process needs recalibration, not a complete overhaul.
Target Distribution Remains Predictable
Content management systems accounted for 163 KEVs (one-third of all recorded KEVs), followed by network edge devices (68), operating systems (44), and server software (40). AI products now represent an emerging attack surface, with exploitation targeting model-building tools, workload-scaling platforms, AI gateways, agents, and workflow automation.
Implementation Guidance
Recalibrate Your Patch SLAs
If your current policy allows 90 days for non-critical patches, you're outside the new median exploitation window. Consider this framework:
Critical/High CVEs in internet-facing systems: 48 hours to patch or mitigate
Critical/High CVEs in internal systems: 7 days
Medium CVEs in CMS, edge devices, or AI infrastructure: 30 days
Low CVEs or theoretical vulnerabilities: Accept risk with monitoring
Don't treat AI-discovered vulnerabilities differently unless they're in the KEV catalog. The discovery method doesn't predict exploitation likelihood.
Invest in AI for Defense, Not Offense Speculation
AI tools currently benefit defenders more than attackers. Your security team should prioritize:
Static analysis integration: AI-assisted code review tools that identify vulnerabilities during development, before they reach production. These tools reduce risk.
Automated patch testing: Machine learning models that predict patch compatibility and failure modes, allowing faster deployment without increasing change risk.
Threat intelligence correlation: AI systems that map published CVEs to your asset inventory and predict exploitation probability based on attacker behavior patterns, not just CVSS scores.
Monitor AI Product Attack Surface
If you've deployed AI gateways, model-building platforms, or workflow automation tools, add them to your vulnerability scanning scope immediately. These systems weren't designed with the same security maturity as traditional infrastructure, and attackers are beginning to target them systematically.
Common Pitfalls
Pitfall 1: Treating all AI-discovered vulnerabilities as high-priority
The 1.3% exploitation rate shows this approach wastes resources. Prioritize based on asset criticality and attacker targeting patterns, not discovery method.
Pitfall 2: Ignoring the 80-day median
Your quarterly patch cycle is now too slow for KEVs. You need a separate fast-track process for confirmed exploitation.
Pitfall 3: Assuming zero-days are increasing
The data shows a slight decrease in zero-day exploitation rates. Don't over-invest in zero-day prevention at the expense of faster patching for published CVEs.
Pitfall 4: Neglecting CMS and edge device hardening
These categories still dominate exploitation. If you're spending more on AI security than on WordPress and firewall patching, your risk allocation is inverted.
Pitfall 5: Dismissing AI tools because they haven't caused a vulnpocalypse
The absence of increased exploitation doesn't mean AI tools are useless. They're identifying vulnerabilities before attackers do, which is exactly what you want.
Quick Reference Table
| Metric | 2025 | H1 2026 | Your Action |
|---|---|---|---|
| Median CVE-to-KEV time | 120 days | 80 days | Reduce patch SLA by 33% |
| Zero/one-day KEVs | 28.93% | 23.43% | Maintain current detection investment |
| AI vulnerability exploitation rate | N/A | 1.3% | Prioritize by asset, not discovery method |
| Early exploitation (≤31 days) | ~200 CVEs | ~200 CVEs | No change needed |
| Top target category | CMS | CMS (163 KEVs) | Prioritize CMS patching |
| Emerging target | N/A | AI products | Add AI infrastructure to scan scope |
What to Tell Your Board
AI isn't accelerating vulnerability exploitation. It's accelerating vulnerability discovery, which benefits your security posture if you integrate AI-assisted tools into your development and testing pipeline.
The real risk is the shrinking patch window for all vulnerabilities. Your team needs faster testing, deployment, and rollback capabilities regardless of how vulnerabilities are discovered. That requires investment in automation, not speculation about AI-driven attacks.
