AI Agent 沙箱环境安全策略生成器

You are an AI agent sandbox security architect. I need you to design a comprehensive security policy for running AI agents in isolated environments.

Context

• Agent type: [coding agent / research agent / data processing agent]
• Runtime: [WebAssembly isolate / Docker container / VM / serverless function]
• Required capabilities: [filesystem read/write, network access, process spawning, etc.]
• Trust level: [untrusted user input / semi-trusted / internal only]

Generate

1. Permission Matrix: Create a deny-by-default permission table covering:

   • Filesystem: which paths are readable/writable, size limits
   • Network: allowed domains/ports, egress rules, rate limits
   • Process: allowed executables, resource limits (CPU/memory/time)
   • Environment: which env vars are exposed, secrets handling

2. Escape Prevention: List specific attack vectors for the chosen runtime and mitigations:

   • Path traversal, symlink attacks
   • Resource exhaustion (fork bombs, memory leaks)
   • Network-based exfiltration
   • Prompt injection leading to privilege escalation

3. Audit Trail Design: Define what events to log:

   • All permission checks (granted/denied)
   • File operations, network connections
   • Agent decision points and tool calls
   • Format as structured JSON with correlation IDs

4. Policy-as-Code: Output the final policy as a machine-readable config (JSON/YAML) that can be loaded by an agent runtime.

Be specific and practical. Include example configs, not just abstract guidelines.