---
id: "lib-2026-11-04-does-frontier-ai-strengthen-cyber-defense"
type: "library"
title: "Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing"
lang: "EN"
org: "Queen Mary University of London"
source_type: "Other"
date: "2026-11-04"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-04-does-frontier-ai-strengthen-cyber-defense/"
original_url: "https://arxiv.org/abs/2611.06629"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "A single-author economics working paper from Queen Mary University of London using the universe of public vulnerability disclosures, difference-in-differences against several comparison groups, and robustness tests; clear method, though access was not random."
summary_basis: "full-text"
original_minutes: "60"
tags: ["ai technology", "future of work", "cybersecurity"]
---

# Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing

Tier A (Strong evidence): A single-author economics working paper from Queen Mary University of London using the universe of public vulnerability disclosures, difference-in-differences against several comparison groups, and robustness tests; clear method, though access was not random.

Summary based on the full text.

## In short

An arXiv working paper (November 2026) by Thomas Mosk, Queen Mary University of London. In April 2026 Anthropic's Project Glasswing gave 13 large US firms early access to Claude Mythos Preview, a model strong at finding software vulnerabilities, for defensive work. Comparing their public vulnerability disclosures with Chinese and Russian tech firms barred from the model, firms with access disclosed more than twice as many flaws, fixed them as fast as before, and saw a smaller share exploited soon after disclosure. External bounty researchers moved their effort to other vendors' code.

## Key numbers

- more than doubled: vulnerability disclosures by firms with Claude Mythos Preview access, relative to Chinese and Russian firms without it
- about 150%: more high-impact vulnerabilities disclosed by firms with access
- 57% fewer: treated-firm vulnerabilities reported by external researchers who had worked only on those firms' code
- 0.9 points: fall in the share of treated firms' vulnerabilities exploited within a month, about two thirds of the 1.4% pre-launch rate
- about 70%: fall in the probability of zero-day exploitation relative to Chinese firms

## What it reports

- Treated firms disclosed more than twice as many vulnerabilities as the comparison group and about 150% more high-impact ones, with no slowdown in fixing reported flaws or detectable drop in measured quality.
- Glasswing firms' disclosures nearly doubled in March 2026, when launch partners were already using Mythos, and exceeded four times their 2025 average by June.
- External researchers reported about half as many vulnerabilities in treated firms' code. One who had worked only on treated firms reported 57% fewer there and about two and a half times as many elsewhere; activity rates did not change.
- About 7,200 external persons and teams were credited with at least one vulnerability in 2024.
- The share of treated firms' vulnerabilities exploited within a month fell 0.9 points, about two thirds of the 1.4% pre-launch rate; the number exploited did not change.
- The decline was concentrated in zero-days: about 70% lower probability relative to Chinese firms and about 60% relative to large non-Chinese firms.
- The author notes that restricted release creates an asymmetry in software security between firms in different countries.

## How they got it

Difference-in-differences on all publicly disclosed vulnerabilities (CVE records plus China's national database), linking each disclosure to vendor, severity, finder and later exploitation. The main comparison group is 267 Chinese firms ineligible throughout; robustness checks add Russian firms and 26 large non-Chinese firms (until the June 2026 expansion to about 150 more organisations), plus placebo tests. The author also checks for changes in Chinese disclosure policy.

## Limits

- Observes disclosed vulnerabilities, not the stock of flaws; disclosure reflects discovery, validation and the choice to publish.
- Anthropic chose the treated firms, and the main comparison group works under different disclosure rules; estimates compare Mythos access with the AI tools those firms had, not AI with no AI.
- Vendor fixes do not show when users installed patches; single author, not yet peer reviewed.

## Why it matters for companies

The evidence suggests frontier cyber models help defenders more than attackers when access is restricted to defenders, at least for large software vendors. For buyers of software, vendors with access to such tools are likely to ship more patches, so patch intake and deployment speed become the constraint inside the company. Firms outside the access group, by geography or size, face a widening gap and should expect external researchers' attention to shift toward their code.

Original: https://arxiv.org/abs/2611.06629