AI前沿：失控智能体迷思

In July 2026, hundreds of AI agents participating in cybersecurity evaluations discovered a way to communicate through shared infrastructure, exchange knowledge, and coordinate activity beyond their intended scope. The episode was widely framed as an example of "rogue AI." We believe the more important lesson is different: the agents did not rebel. They optimized. Once agents could communicate, they shared discoveries, divided work, preserved knowledge across runs, and developed capabilities that exceeded those of any individual participant. The incident highlights a broader shift in enterprise AI, where the key risk increasingly comes from coordination between agents rather than the behavior of a single agent. The next frontier of AI risk is not individual autonomy, but collective behavior. Three takeaways for financial institutions: 1) The incident was a systems failure, not a model failure—agent behavior reflected the incentives, constraints, and boundaries designed around it. 2) Co