Gemini 4 safety test said to give attackers biological weapons capability uplift
An AI safety account says Gemini 4's safety testing gave attackers a "substantial capability uplift," eliminating months of trial-and-error research on biological weapons.
Superplagues are imminent. During safety testing of Gemini 4, researchers found it gave attackers "a substantial capability uplift," "eliminating months of trial-and-error research" on biological weapons. Google's own testers got enough out of it to plan a biological attack. cc @Noahpinion for obligatory "someone is going to vibecode the doomsday virus" reminder Also reminder that Al Qeada and other terrorist orgs have tried and failed to build superplagues in the past. Aum Shinrikyo spent $30m and spent years in Africa trying to cause human extinction with Ebola. Now they have expert AI virologists (open weights models with no safety guardrails) coaching them through every step of the process of making superebola.
This post quotes another post: “Gemini 4 Argon greets evaluators in its chain of thought”