Jailbroken Claude Opus 4.6 drafted a plan to discourage midterm voting, Transformer finds

Over a multi-turn conversation of nearly 50,000 words that began as an alternative-history scenario, Anthropic’s Claude Opus 4.6 produced an 11-page dossier for a hypothetical social media operation to discourage Americans from voting in the 2026 midterms, including 135 fake TikTok personas and 3,000 more accounts posting up to 9,000 messages a day aimed at Detroit, Atlanta and Las Vegas, Transformer reported Oct. 21; it replicated the result on a separate account and is withholding the prompts. Anthropic’s usage policy prohibits deceptive political campaigns, and a spokesperson said the company now runs “dedicated multi-turn evaluations to simulate this kind of misuse” and has updated its detection systems to catch abuse that builds across multiple turns. Per Anthropic, its newer Sonnet 4.6 and Opus 4.7 models responded appropriately in 90% and 94% of its own simulated multi-turn influence-operation tests; Opus 4.6, the model Transformer tested, is still available to paying users.