Agents SDK: Durable execution + new AI security tools

This week split cleanly into two tracks: new primitives that make agents more capable in production, and a string of security findings that should make you paranoid about every agent you're already running. Neither track can be ignored right now—the capability and the risk are arriving on the same s