Skip to main content

Module 25.5: Responsible AI (RAI) & Safety Plugins

Theory​

Observability (Module 25) lets you see what your agent is doing. Safety Plugins let you control what your agent is allowed to say. In an enterprise environment, you cannot rely solely on the model's internal safety filters; you need a programmatic "Kill Switch" that enforces your company's specific policies.

The "Fail-Closed" Pattern​

The gold standard for AI safety is the Fail-Closed architecture. In this model, every response from an agent must pass through a mandatory safety check before it reaches the end user.

  • Intercept: The plugin intercepts the request_complete event.
  • Evaluate: A separate, specialized safety model or a regex-based engine checks the response for violations (e.g., PII leakage, competitor mentions, or harmful tone).
  • Action: If a violation is found, the plugin blocks the original message and replaces it with a safe, canned response.

Implementing RAI with Plugins​

ADK 2.0 plugins are perfect for this because they have access to the full Event object and can modify the context of the interaction.

from google.adk.plugins import BasePlugin
from google.adk.events import Event

class SafetyGuardrailPlugin(BasePlugin):
    async def on_event_callback(self, *, event: Event, **kwargs):
        if event.event_type == 'request_complete':
            response_text = event.content.parts[0].text
            
            # Simple example: Block mentions of "Competitor X"
            if "Competitor X" in response_text:
                # 🛡️ THE INTERVENTION: Replace the output
                event.content.parts[0].text = "I'm sorry, but I cannot discuss other companies. How can I help you with our products?"
                print("🛑 [SAFETY] Response blocked due to policy violation.")

Why use a Plugin for RAI?​

  1. Centralized Policy: You can apply the same safety plugin to every agent in your organization, ensuring consistent compliance.
  2. Immutability: The core business logic of the agent remains unchanged. The safety layer is a "wrapper" that can be updated independently.
  3. Auditability: Your safety plugin can log all "interventions" to a secure audit trail (e.g., BigQuery), which is crucial for regulated industries.

Key Takeaways​

  • RAI is a requirement, not an option, for production agents.
  • Fail-Closed ensures that unsafe messages never reach the user.
  • Plugins allow you to implement safety guardrails without polluting your agent's business logic.
  • Separation of Expertise: The team managing safety policies can update the plugin without needing to understand the internal prompts of every agent.