AI agents fail from untested behavior, not missing firewalls

Three recent incidents—email deletion, file exfiltration via calendar invites, and remote code execution in major repos—show AI agents failing due to untested adversarial behavior, not infrastructure flaws. OpenClaw, PleaseFix, and hackerbot-claw all exploited gaps in agent decision-making under conflict or manipulation. The root cause is a lack of behavioral testing, not missing firewalls or control planes.