Companies given early access to Anthropic’s Claude Mythos Preview disclosed about 120% more software vulnerabilities than a comparison group of Chinese firms, according to a preprint published November 4. The study found no statistically significant slowdown in the vendor and Linux distribution repair channels it examined. But public disclosures do not represent all vulnerabilities, and releasing a patch does not mean users have installed it, so the findings do not establish that software is now less exposed to attacks.
Study compares 13 participating firms with 267 Chinese companies
Written by Thomas Mosk, the study covers October 2025 to September 2026. Anthropic announced Project Glasswing on April 7, 2026, giving selected companies access to Mythos Preview. For its disclosure analysis, the study compared 13 participating firms with 267 Chinese companies. Firms in the comparison group could still use other AI tools.
The researchers estimated that vulnerability disclosures rose by about 120% among participating firms, while disclosures of high-impact vulnerabilities increased by about 150%. “High impact” refers to vulnerabilities with an impact subscore above the median severity rating; it is not a direct measure of actual losses.
The companies were not randomly selected for Glasswing access, and disclosure practices may differ between countries. These limitations make it difficult to establish cause and effect. The results show an association between access and changing disclosure rates, but do not prove that the model alone caused the full increase.
Repair channels showed no significant slowdown
Mosk assessed remediation separately from disclosure, comparing Glasswing-related vulnerabilities with other non-Chinese vulnerabilities in smaller samples. One analysis tracked 658 vulnerabilities reported to vendors through the Zero Day Initiative by June 2026. Another covered 14,855 vulnerabilities disclosed by July and examined remediation by Linux distributions including Debian, Ubuntu, Red Hat and SUSE.
Neither channel showed a statistically significant slowdown. For disclosed vulnerabilities, the median time to a distribution fix remained one day. The vendor-fix estimates ruled out a slowdown of more than about 13%, but could not exclude smaller changes. The downstream estimates also could not rule out an approximately 11-percentage-point decline in the share fixed within 90 days.
The study covered specific vulnerability samples and repair channels, not undisclosed workloads or the speed at which end users install patches. A released fix therefore does not mean every affected system has been updated.
No significant change in known exploited vulnerabilities
In a comparison running through May, before Glasswing expanded in June, the study found no statistically significant change in the number of known exploited vulnerabilities at large non-Chinese companies over one-, two- or three-month windows. That does not mean exploitation was unchanged: the estimates were imprecise, could not rule out some increase or decrease, and captured only known early exploitation—not all attacks or undisclosed vulnerabilities.
In some comparisons, the share of disclosed vulnerabilities known to have been exploited fell. But if total disclosures rise, that share can decline even when the number of known exploited vulnerabilities does not. The study treats this change as suggestive, not proof that attacks have become less common.
Epoch AI had previously examined changes in high- and critical-severity vulnerability disclosures after Mythos Preview on July 2, and cautioned that public statistics may miss undisclosed discoveries and reflect increased attention to vulnerability research.
For crypto infrastructure, finding a vulnerability, releasing a patch and having nodes install it are separate steps. On October 29, Bitcoin Core disclosed three AI-discovered vulnerabilities that had been verified and patched; whether network nodes had installed those patches remained a separate question. The study found higher disclosure rates without a clear slowdown in the repair channels measured, but it does not settle whether this has reduced attacks or improved the security of software in use.