New hack uses prompt injection to corrupt Gemini’s long-term memory
On Monday, researcher Johann Rehberger demonstrated a new way to override prompt injection defenses Google developers have built into Gemini—specifically, defenses that restrict the invocation of ...