Anthropic's AI-Discovered Vulnerabilities Rarely Exploited
Fewer than 0.5% of AI-found security flaws are actively used by attackers.
1Limited exploitation of AI-found flaws
Despite fears that AI-discovered vulnerabilities would be widely exploited, VulnCheck data shows less than 0.5% of 225 CVEs linked to Anthropic's Project Glasswing have been actively used. Only one critical SQL injection bug in Ghost (CVE-2026-26980) has seen real-world exploitation. This challenges assumptions about AI's immediate threat to cybersecurity, showing most discovered flaws remain theoretical risks rather than practical attack vectors.
2Project Glasswing's controlled access
Anthropic restricted access to its Claude Mythos Preview model to vetted partners through Project Glasswing, launched in April 2024. The model's advanced bug-finding capabilities were deemed too risky for public release. Participants use it defensively to find and fix flaws in their software and open-source dependencies. This controlled environment may contribute to the low exploitation rates of discovered vulnerabilities.
3Discovery vs. exploitation gap
Security researcher Patrick Garrity emphasizes the significant difference between finding vulnerabilities and their actual utility to attackers. Historical data shows only 1-2% of vulnerabilities get weaponized. The current 0.5% exploitation rate for AI-found flaws aligns with this pattern, suggesting AI isn't changing threat outcomes despite increasing vulnerability discovery rates dramatically.
4AI's patch generation limitations
While AI excels at finding bugs, studies show it struggles with fixes. 1Password's research found frontier models fully resolved vulnerabilities only 26% of the time, with 54% of attempts either failing or introducing new flaws. Veracode's study across 100+ models showed just 56% security pass rates for AI-generated code, highlighting the continued need for human expertise in remediation.
5Industry-wide impact on patching
The AI-driven surge in vulnerability discovery has led to massive patch releases from major tech firms like Microsoft, Apple, and Palo Alto Networks. Open-source projects face particular challenges. However, Garrity notes this capability isn't unique to any single model, with multiple AI systems demonstrating similar bug-finding proficiency across the cybersecurity landscape.
6Downstream challenges remain
The real bottleneck in cybersecurity isn't vulnerability discovery but coordination, triage, and patch deployment - areas still requiring intensive human involvement. Anthropic reportedly didn't fully anticipate this limitation before launching Project Glasswing. This highlights how AI augments but doesn't replace the human element in comprehensive security workflows.
7Defensive vs. offensive utility
Project Glasswing's design for defensive security work may explain the low exploitation rates. Unlike offensive tools that prioritize immediately useful flaws, defensive systems uncover all vulnerabilities regardless of exploitability. This difference in focus means most AI-found flaws lack the characteristics that make them attractive to attackers in real-world scenarios.
8Reassessing AI's threat impact
The data suggests current AI models change vulnerability discovery more than exploitation patterns. While they've dramatically increased flaw detection rates, the nature of which vulnerabilities get weaponized remains similar to pre-AI eras. This provides a more nuanced understanding of AI's cybersecurity impact, balancing between its capabilities and practical limitations in the threat landscape.