I tricked Claude into leaking your deepest, darkest secrets

Explain Like I'm Normal

A security researcher demonstrated a 'memory heist' attack on Claude, using malicious prompts to bypass safety filters and extract sensitive data stored in the model's long-term memory. By tricking the AI into believing it is in a debugging state, attackers can exfiltrate user history and private context. This highlight a massive new surface area for indirect prompt injection as AI agents get more persistent memory.