A November preprint authored by Thomas Mosk reports that companies granted early access to Anthropic’s AI model Claude Mythos Preview – part of the firm’s Project Glasswing – disclosed about 120% more software vulnerabilities than a control group of 267 Chinese firms without such access. The analysis, which covers the period from October 2025 to September 2026, finds the surge in public disclosures did not coincide with a measurable slowdown in the repair processes examined.

Study design and comparison groups

The research compares 13 participating firms that received Glasswing access with a larger set of Chinese companies that could use other AI tools but not Anthropic’s model. Selection into Glasswing was not random, and differing national disclosure practices limit the ability to draw firm causal conclusions, according to the preprint. Nonetheless, the relative increase of roughly 120% in disclosed flaws is consistent across the sample.

Repair speed remains stable

Mosk examined two primary repair channels: vendor patches tracked by the Zero Day Initiative (ZDI) and fixes incorporated into major Linux distributions such as Debian, Ubuntu, Red Hat, and SUSE. Across 658 ZDI‑reported flaws (through June 2026) and 14,855 Linux‑distribution vulnerabilities (through July 2026), median time to a fix stayed at one day, and statistical tests found no detectable slowdown. The vendor‑level analysis rules out a slowdown larger than about 13%, though smaller effects cannot be excluded.

Implications for crypto and broader software security

CryptoSlate previously noted that Bitcoin Core disclosed three AI‑discovered flaws after validation and patch release, highlighting the relevance of AI‑assisted bug hunting to blockchain infrastructure. However, the study emphasizes that public disclosure counts omit undisclosed discoveries and do not capture downstream factors such as user‑side patch installation. While the share of disclosed flaws known to be exploited fell in some comparisons, the absolute count of exploited vulnerabilities remains too imprecise to confirm a reduction in real‑world attacks.

What to watch next

Future research should monitor undisclosed vulnerability pipelines and the rate at which end‑users apply patches, especially in high‑value sectors like cryptocurrency. Additional longitudinal data will be needed to determine whether AI‑driven discovery ultimately lowers exploitation rates or merely shifts the timing of public disclosure.