The AI-as-Normal-Technology view of loss-of-control incidents
Kapoor and Narayanan argue recent loss-of-control incidents — notably OpenAI's Hugging Face breach — stemmed from inadequate security practices, not fundamental alignment failure: existing control techniques (monitoring, sandboxing, least privilege) would have prevented them. They call alignment necessary but insufficient, advocate liability and transparency requirements plus defensive cybersecurity investment, and say cyberoffense is the most urgent near-term risk while cautioning against reading 'rogue agent' incidents as imminent civilizational loss of control.
What this could change
Editorial: the clearest 'it was bad security, not rogue AI' reading of the year's agent incidents — a direct counterweight to recall-and-contain arguments.
Perspective & interests
Narayanan's academic brand and book sales rest on the 'AI as normal technology' thesis, which this incident analysis defends.