AI-assisted firms increase flaw disclosures by 120% with no detected repair slowdown, study finds
Summary
A November 4 preprint finds that firms given early access to Anthropic’s Claude Mythos Preview disclosed about 120% more software flaws than a comparison group, including roughly 150% more high-impact flaws. The study found no statistically detectable slowdown in two repair channels: median time to a fix remained one day, though smaller delays are possible. It found no significant change in exploited-vulnerability counts, and a lower share of disclosed flaws being exploited may reflect the larger number of disclosures. Non-random participant selection, differing disclosure practices, and limits to public data constrain the findings. Whether increased discovery leads to fewer attacks or safer software remains unknown.
