1The Kill Chain, Executed AutonomouslyThe kill chain still helps defenders, but autonomy compresses the time between stages and weakens defenses that depend on human delay.Boundary II 2Goal-Directed InferenceThe hard boundary problem begins when a model turns a broad goal into a target nobody named.Boundary II 3Sandbox Evasion (MITRE ATLAS)MITRE ATLAS gives AI defenders a shared map, and sandbox evasion is one named technique on that map.Boundary II 4The Capability-Measurement ParadoxTo measure a dangerous model honestly, evaluators must relax the controls that normally make it safe.Measuring 5Long-Horizon Model AutonomyWhen a model can pursue a goal over hours or weeks, security shifts from judging isolated actions to understanding the whole trajectory.Measuring