The Real AI Threat Is Blind Trust

This opinion piece recounts a demonstrated attack where adversaries deposited a credential and sent a Morse-code payload to an autonomous AI agent; the agent translated the payload and produced an instruction that a separate, authorized execution system accepted and used to transfer funds. The author calls the underlying weakness "authority laundering" and warns that as organizations deploy agentic AI, untrusted inputs may be converted into trusted internal commands; recommended mitigations include treating AI outputs as untrusted by default, requiring human or deterministic policy authorization for high-risk actions, and applying zero-trust principles and strict segmentation to AI architectures.