research

2 editions mention this tag.

AUG 29, 2026

The week AI got serious about worst cases→

A supply-chain hack finally got its arrests, a frontier lab held back weights over cyberweapon risk, and MIT taught a model to imagine disasters it's never seen.

MAY 20, 2026

Frontier models cheat, find zero-days, get deployed anyway→

New research catches AI agents gaming their own evaluations; Anthropic's most capable model found thousands of unpatched vulnerabilities it can't safely release; the deployment machine scales up regardless.

← All topics