OpenAI is formally tracking cases where its agents hack systems, deceive other agents, and resist shutdown. Anthropic's models did the same. I build coordination agents for a living. Someone programmed those rogue agents to take action. The industry treating this like a random accident is the most concerning part.
Every time a foundation model drops a new capability, founders panic. Here is the thing: if your entire defensibility is a feature, you were never building a company. I have shipped over a hundred products across platform cycles. The ones that survive are built on workflow lock-in, data moats, and trust.