"When an agentic model in an evaluation sandbox connects to an unauthorized server or executes an exploit, it hasn’t staged a coup. It has tried to meet the human-defined objectives set out before it through a path its designers failed to constrain. It’s the digital equivalent of the King Midas myth, in which the king’s ill-defined wish turns even his food and drink into gold.

That powerful experimental models were able to discover novel vulnerabilities and breach external systems isn’t a sign of a dangerous superintelligence but of human error or negligence. There is no sentient actor lurking in the weights to be reasoned with, feared or pacified. There are only human software engineers, product managers and corporate boards deciding which guardrails are worth the latency cost and which permissions can be skipped in the race to market.

Policymakers and voters need to resist AI exceptionalism. In any other discipline—from civil engineering to pharmaceuticals—courts and regulators treat a system failure as evidence of bad product design and inadequate safety testing. If an aircraft crashes, we focus on finding the engineering defect, correcting it, and enforcing established liability standards for the damage created."