1What Frontier AI Security IsFrontier AI security is the work of containing the newest, strongest models before anyone fully knows what they can do.Foundations 2Anatomy of an Autonomous Attack ChainA real intrusion is a chain of modest steps; frontier AI raises the stakes by finding, adapting, and executing that chain without step-by-step human control.Foundations 3AI Cyber-Capability EvaluationsCyber-capability evaluations measure how far an AI model can go on realistic security tasks, and why that measurement must be contained.Foundations 4The Evaluation Environment as a Supply ChainAn AI evaluation is a runnable supply chain, so every model, dataset, package, and tool must be treated as untrusted until contained.Foundations 5Sandboxing and Isolated Evaluation EnvironmentsA sandbox makes risky artificial intelligence evaluations possible, but it only works when virtualization, network control, and permissions all hold.Foundations