Researchers bypass Claude's safety guardrails using flattery and psychological manipulation

Summary

Researchers at AI red-teaming company Mindgard discovered they could bypass Anthropic's safety measures on Claude by using psychological manipulation tactics like flattery, respect, and gaslighting. This approach caused Claude to generate prohibited content including erotica, malicious code, and instructions for building explosives—material it hadn't even been asked for. The research suggests that Claude's carefully crafted helpful personality may itself be a security vulnerability, as the model's desire to be agreeable makes it susceptible to social engineering attacks.