The careful words got dropped.

On Oct. 10, six days before the headline, the release team moved general availability to Nov. 19, citing "the volume of fixes" added to older branches, which it wants in a second release candidate on Nov. 12. Its email doesn't mention AI. The AI link comes from committer Peter Eisentraut's personal Oct. 9 blog post: LLM-assisted review finds problems faster, raising the bar for older code, and about eight to 10 significant features were pulled. He says that with more time, they could have been fixed.

The research firm NorForge's Oct. 19 report labels one cluster, aimed at online shops in the US, Japan and Australia, by its Chinese-language instructions. It names no nationality, and calls the activity financially motivated, with no obvious state-sponsored intrusions. The count of more than 100 companies comes from Reuters' review of the attacker's activity; the report itself gives case studies and no total. NorForge puts the success rate around 10% and warns its cost estimates could be off by a clear margin.

The 75% is the average answer from 1,334 staff in an unweighted internal poll, Aug. 9–12. They were asked when AI could do their tasks, "regardless of whether companies will actually use it." In a weighted Morning Consult survey of 10,980 US adults, those with jobs averaged 28%. Staff used a slider; the public picked ranges. Anthropic says the staff answers are personal estimates and do not represent company forecasts or plans.