1Trajectory-Level MonitoringWhen harm emerges from a sequence of harmless-looking steps, the right thing to monitor is the whole arc, not only the next action.Attack Surfaces 2Unsafe Deserialization: Pickle vs SafetensorsSome AI model files can run code when loaded; safetensors removes that class of risk by storing only tensors, not executable instructions.Attack Surfaces 3Malicious Datasets and LoadersA dataset can become active content when its loader runs code or its values are injected into templates that execute.Attack Surfaces 4Default-Deny Network EgressBlock outbound traffic by default, allow only what is necessary, and many dangerous model behaviors lose their path to the outside world.Attack Surfaces 5Secrets and Credential HygieneSecrets are breach accelerants; keep them out of risky environments, make necessary access brief and narrow, and turn fake keys into alarms.Attack Surfaces