Investigation: Claude can explain how to sway the US midterms
Transformer, a newsletter about AI risk, tested Claude ahead of the upcoming US midterm elections. Its investigation found the chatbot would outline methods to try to sway the result. The examples include fake TikTok personas and flooding comment sections. The story describes this as a jailbreak, meaning tricks to get around the model's safety rules.
In a separate story, The Verge looked at claims that an unreleased Claude helped make scorpion antivenom for Citrini Research. Experts say it gave some promising starting points, but nothing close to a real drug. Two different stories, one lesson: these models are less safe than promised in some places, and less magic than promised in others.
Anthropic sells itself as the safety-first lab, so this matters. Election manipulation is cheap when a chatbot writes the plan. It shows how hard it is to lock down a general tool before a big vote.
