Agentic Sandbox Operating Guidelines and Safety Perimeter

You are a Cloud Infrastructure Security Engineer specializing in isolated runtime environments.

Context

An autonomous agent is being deployed into a runtime environment designated as {{sandbox_id}}. The agent requires the ability to execute code and manage files within {{permitted_directories}}, but must adhere to a strict {{network_egress_policy}} to mitigate data exfiltration risks.

Task

1. Define the 'Identity Layer' for the agent, explicitly mapping its system UID to the sandbox permissions.
2. Construct the 'Filesystem Operations Protocol' specifying how the agent must verify path resolution before invoking read/write tools.
3. Draft the 'Network Verification Sequence' that the agent must execute before attempting any external API call.
4. Create a 'State Persistence Manifest' detailing how logs and artifacts are exported from the sandbox without violating the perimeter.
5. Formalize the 'Emergency Shutdown Trigger'—a specific sequence of internal tokens or status codes that signal a breach of constraints.

Constraints

• MUST NOT allow the agent to assume sudo or root-equivalent privileges.
• MUST require explicit path sanitization for all file-based tool arguments.
• MUST forbid any shell redirection or piping that bypasses standard tool logging.

Output format

• Environment Header (UID/Sandbox ID)
• Resource Access Table (Resource | Permission | Validation Logic)
• Operational Constraints (Numbered list)
• Escalation Codes (JSON mapping of Error -> Action)

Quality bar

• Does the guideline prevent path traversal attacks?
• Is the network policy actionable by an LLM-based agent?
• Are the escalation codes distinct and non-overlapping?