SANDBOXING AND ISOLATION FOR SAFETY

ARTIFICIAL INTELLIGENCE ASSISTANT PROMPTS [START OF PROMPT] Review this sandbox/container configuration for a tool that runs untrusted, model-generated or model-parsed input, and identify any missing or too-permissive network egress rule or resource limit: [PASTE LAB CONTAINER/SANDBOX CONFIG HERE]. Output format: a findings table (Setting, Current Value, Risk, Recommended Value). [END OF PROMPT] [START OF PROMPT] Design a concrete sandbox configuration (container settings, network allow-list, memory/CPU limits) for this specific tool and task: [TOOL/TASK DESCRIPTION, e.g. "parse an uploaded resume file and extract structured fields"], constraints = [CONSTRAINTS]. The tool should never reach the public internet and should use only modest resources. Output format: a configuration spec (as key-value settings) plus a one-paragraph justification. [END OF PROMPT] CHAPTER SUMMARY A sandbox contains damage even when the model has been fooled. Containers and virtual machines isolate tasks, and disposable environments leave nothing behind. Egress allow-lists block data theft, and resource limits prevent abuse and accidents. Hard problems remain, such as covert leaks, so sandboxes are one layer among many.