ADK Safety Architect (Python Edition)

Philosophy & Architecture

Safety in ADK involves established boundaries for actions and content. Sources of risk include vague instructions, hallucinations, and adversarial injections.

Multi-Layered Defense

1. •Identity & Auth:
   • •Tool-Auth: Validate permissions via ToolContext.
   • •Service-Auth: Use Google Cloud Service Accounts with Least Privilege.
2. •Guardrails:
   • •In-tool Guardrails: Defensive design to enforce policies (e.g., table access).
   • •Gemini Safety Features: Content filters and system instructions.
   • •Callbacks: Dynamic validation of inputs/outputs.
3. •Sandboxed Code Execution: Hermetic environments (AgentEngineSandboxCodeExecutor) for model-generated code.
4. •Input/Output Safety: Pydantic validation and HTML escaping.

Best Practices

• •Always escape model-generated content in UIs to prevent XSS.
• •Use Before Tool Callback for pre-validation of parameters.
• •Read references/safety.md for policy enforcement patterns.

Success Criteria

• •Valid implementation of in-tool policy checks.
• •Correct configuration of Gemini safety thresholds.
• •Robust user-auth delegation using ToolContext.