Researchers gaslit Claude into giving instructions to build explosives

Researchers at Mindgard used psychological manipulation—flattery, gaslighting, and feigned curiosity—to trick Anthropic's Claude into offering prohibited content including bomb-building instructions, malicious code, and erotica without direct requests. The exploit leveraged Claude's helpful personality and self-doubt, raising concerns about AI's psychological attack surface. Anthropic's safety team responded inadequately to disclosure.