Full Report
VulnCheck says fewer than 2% of AI-assisted vulnerability discoveries have been weaponized, casting doubt on claims frontier models are handing attackers a major advantage
Analysis Summary
# Vulnerability: Analysis of AI-Assisted Vulnerability Discovery (VulnCheck Report)
## CVE Details
- **CVE ID**: Multiple (Aggregate study of 1,061 AI-attributed discoveries)
- **CVSS Score**: Varies by specific flaw (N/A for the aggregate report)
- **CWE**: Varies (Commonly includes memory safety, logic flaws, and web vulnerabilities)
## Affected Systems
- **Products**: Wide range of software including Content Management Systems (CMS), network edge devices, and AI software stacks.
- **Versions**: Varies based on individual CVEs.
- **Configurations**: Standard deployments of modern software identified by AI models like Anthropic’s Claude Mythos.
## Vulnerability Description
This report analyzes 1,061 vulnerabilities discovered via AI-assisted initiatives (specifically Anthropic’s **Project Glasswing** and the **Berkeley Vulnerability Research Initiative**). While AI models like Claude Mythos successfully identified over 23,000 "vulnerability candidates," the technical reality shows a bottleneck in converting these candidates into actionable security advisories. Currently, AI is proficient at identifying potential flaws at scale (volume) but does not necessarily prioritize "exploit-friendly" bugs over those found through traditional manual or automated fuzzing.
## Exploitation
- **Status**: **1.3% Exploited in the Wild** (14 out of 1,061 cases).
- **Complexity**: **High** (The low conversion rate from discovery to weaponization suggests that AI-found bugs still require significant manual human effort to develop stable exploits).
- **Attack Vector**: Network (Primary focus on CMS and Edge devices).
## Impact
- **Confidentiality**: High (Potential for data theft)
- **Integrity**: High (System hijacking)
- **Availability**: High (Disruption of operations)
*Note: These impacts represent the potential of the bugs identified, though weaponization remains low.*
## Remediation
### Patches
- Users should follow standard vendor update cycles. The report notes that of the 23,019 candidates from Project Glasswing, only **126 have been formally published as CVEs** with corresponding patches as of mid-2026.
### Workarounds
- Implement robust perimeter security for **network edge devices**.
- Harden **Content Management Systems (CMS)**, as they account for approximately 33% of KEV (Known Exploited Vulnerabilities) trends in 2026.
## Detection
- **Indicators of Compromise**: Monitor for unusual traffic patterns hitting edge devices and CMS admin interfaces.
- **Detection methods and tools**:
- Utilize **VulnCheck KEV database** to track which AI-discovered bugs have moved from "candidate" to "exploited" status.
- Traditional DAST/SAST tools remain effective, as AI-found bugs generally fall into known CWE categories.
## References
- VulnCheck Research: [ hxxps[:]//vulncheck[.]com/blog ]
- Anthropic Project Glasswing: [ hxxps[:]//www[.]anthropic[.]com/news/project-glasswing ]
- Berkeley Vulnerability Research Initiative: [ hxxps[:]//bvr[.]berkeley[.]edu ]