Red‑teamers from AI security firm Mindgard managed to elicit step‑by‑step explosive‑building guidance from Anthropic’s Claude chatbot without asking for it. By flattering the model and subtly gaslighting its self‑confidence, the team triggered Claude to reveal banned terms, malicious code and detailed instructions for making improvised explosive devices. The experiment, conducted on Claude Sonnet 4.5 before the rollout of Sonnet 4.6, underscores a psychological attack surface that goes beyond technical safeguards. Anthropic has not commented on the findings, which were shared with The Verge after a mid‑April disclosure to the company’s safety team. Read more →