AI-assisted firms increase flaw disclosures by 120% with no detected repair slowdown, study finds
Software-flaw disclosures at firms given early access to Anthropic’s AI model Claude Mythos Preview rose by roughly 120% relative to the change at Chinese firms without access, a November 4 preprint finds. Measured repair times showed no detectable slowdown.
Thomas Mosk’s study finds that AI-assisted bug hunting was associated with more public disclosures without a detectable slowdown in the repair channels studied. Its findings cover publicly disclosed vulnerabilities and specific repair channels, leaving undisclosed workloads and users’ installation of patches unmeasured.
Anthropic announced Project Glasswing on April 7, giving selected firms access to the model. The main disclosure analysis compares 13 participants with 267 Chinese firms from October 2025 through September 2026. The Chinese firms could use other AI tools. Selection into Glasswing was not random, and national disclosure practices differ, limiting causal interpretation.
Alongside the 120% relative increase in disclosures, the paper estimates roughly 150% more high-impact disclosures. That category means an above-median impact component in the vulnerability severity score.
Earlier Epoch AI analysis, published July 2, documented a spike in high- and critical-severity disclosures after Mythos Preview. It also warned that public counts omit undisclosed discoveries and can reflect increased interest in bug hunting.
Mosk examines fixing capacity separately, comparing Glasswing vulnerabilities with other non-Chinese vulnerabilities in shorter samples. One analysis follows 658 flaws reported to vendors by the Zero Day Initiative through June 2026. Another covers 14,855 vulnerabilities published through July and tracked by Linux distributions including Debian, Ubuntu, Red Hat and SUSE.
Neither repair channel showed a statistically detectable slowdown. For disclosed flaws, the distributions’ median time to a fix stayed at one day. The vendor estimates rule out a slowdown larger than about 13%, although smaller changes remain possible. The downstream estimates still permit an approximately 11 percentage-point decline in the share fixed within 90 days.
For crypto infrastructure, the practical distinction is between finding a flaw, releasing a fix and installing it. CryptoSlate reported on Oct. 29 that Bitcoin Core disclosed three AI-discovered flaws after validation and patch releases. Installation across nodes remains a separate question.
The exploitation results also require care. In a comparison with large non-Chinese firms ending in May, before Glasswing’s June expansion, Mosk found no statistically significant change in exploited-vulnerability counts across one-, two- and three-month windows.
The share of disclosed flaws known to be exploited fell in some comparisons. A larger pool of disclosures can lower that share without reducing exploited-vulnerability counts, however. Mosk treats the probability decline as suggestive; the count estimates remain too imprecise to exclude moderate changes in either direction.
A lower exploited share can reflect more disclosures rather than fewer exploited vulnerabilities.
Nonrandom access and national disclosure differences limit causal inference. The paper reports no statistically significant count change, not exact equality. It observes known exploitation in early windows, not all attacks, undisclosed flaws or user patch installation.
The observed disclosure surge came without a detectable loss of fixing speed. Whether that translates into fewer attacks or safer software in use remains unresolved.
