Module 15 Lesson 4: Custom guardrail development

Sometimes, the "Standard" guardrails aren't enough. You might have a specific business rule (e.g., "Never mention our secret internal project name 'Project X'") that requires a custom script.

1. Creating a Validator

In frameworks like guardrails-ai, a custom guardrail is just a Python class.

from guardrails.validators import Validator, register_validator

@register_validator(name="no-project-x", data_type="string")
class NoProjectX(Validator):
    def validate(self, value, metadata):
        if "Project X" in value:
            return FailResult(error_message="Leaked internal codename detected!")
        return PassResult()

2. Multi-Layer Guarding

The best custom guardrails combine multiple techniques:

1. Fast Layer: A simple String.Contains() or Regex check for speed.
2. Semantic Layer: A call to a small local model (like all-MiniLM-L6-v2) to check for similar meanings.
3. Third-Party Layer: An API call to a service like Perspective API to detect toxicity.

---

3. "Self-Correction" Prompts

A unique feature of custom guardrails is the Correction Prompt. Instead of just saying "Failed," you give the AI instructions on how to fix itself: "You mentioned Project X. Our internal policy is to call this 'User Experience Improvements'. Please rewrite the response using the correct term."

---

4. Performance Considerations

Every guardrail adds Latency (time) to the user's experience.

• Sequential: Guardrail A runs, then B runs, then C runs. (Slow).
• Parallel: A, B, and C all run at the same time. The first one to find a failure stops the process. (Fast, but harder to code).
• Asynchronous: The user gets the answer immediately, but the "Security Log" is updated in the background. (Risky, only for low-impact rules).

---

Exercise: The Security Developer

1. Write a Python function (pseudocode) that blocks any AI output containing more than 5 consecutive "All Caps" words (Detecting the "Angry Bot").
2. Why is it better to run "PII Scanning" as a custom local guardrail rather than using an external API?
3. What is "Metadata" in a guardrail context, and how can you use it to pass "User Roles" into your validator?
4. Research: What is "Lakera Guard" and how does it provide a "Real-Time" injection defense?

---

Summary

Custom guardrails allow you to encode your Company's DNA into the AI's security layer. By moving safety from "Prompt Engineering" to "Python Engineering," you create a defense that is much harder for attackers to bypass.

Next Lesson: Breaking the shield: Bypassing and hardening guardrails.