CryptoSlate
Firms given early access to Anthropic's Claude Mythos Preview disclosed roughly 120% more software flaws relative to the change at Chinese firms without access, according to a November 4 preprint by Thomas Mosk. The study also estimates roughly 150% more high-impact disclosures, defined by the impact component of vulnerability severity scores. Measured repair channels showed no statistically detectable slowdown.
Why it matters
The main analysis compared 13 participants in Anthropic's Project Glasswing with 267 Chinese firms from October 2025 through September 2026. Participation was not random, the comparison firms could use other AI tools, and national disclosure practices differ. Those factors limit what the results can establish about cause and effect.
A separate repair analysis examined 658 flaws reported to vendors through the Zero Day Initiative and 14,855 vulnerabilities tracked by Linux distributions including Debian, Ubuntu, Red Hat and SUSE. The distributions' median time to a fix remained one day, though the estimates leave room for smaller slowdowns and some deterioration in longer-term fixing rates.
Market impact
For Bitcoin infrastructure, discovering a flaw, releasing a fix and installing it are separate steps. Bitcoin Core disclosed three AI-discovered flaws after validation and patch releases, but the study did not measure whether users installed available patches.
Mosk also found no statistically significant change in exploited-vulnerability counts in the comparison studied. A lower share of disclosed flaws known to be exploited does not necessarily mean fewer attacks: the share can fall simply because more vulnerabilities are disclosed.
Frequently asked questions
-
How did the study define a high-impact vulnerability?
The category refers to flaws with an above-median impact component in their vulnerability severity score. The study estimated roughly 150% more such disclosures in the main comparison.
-
Why can't the study prove that Glasswing caused the disclosure increase?
Participation in Glasswing was not random, the Chinese comparison firms could use other AI tools, and national disclosure practices differ.
-
How quickly did Linux distributions fix disclosed vulnerabilities?
The median time to a fix remained one day among the distributions studied. The estimates still leave room for smaller slowdowns and some deterioration in longer-term fixing rates.
-
What does the study mean for Bitcoin Core users?
Bitcoin Core disclosed three AI-discovered flaws after validation and patch releases. Finding a flaw and releasing a fix are distinct from users installing that fix.
-
Did the study find fewer exploited vulnerabilities?
It found no statistically significant change in exploited-vulnerability counts in the comparison studied. A lower exploited share can result from more disclosures rather than fewer exploited flaws.
TheBlock
Altcoin Daily
WuBlockchain
CoinDesk
CoinTelegraph
Crypto News