Four research teams broke AI agents four ways in ten days, from Claude for Chrome to poisoned memory. The AI agent security gap, explained.