Skip to main content

​
Overview

A strong guardrail program starts with the right provider and probe mix for your domain, risk profile, and latency targets.

​
Selection Framework

​
1) Define Risk Domains

Examples:
  • Prompt injection and jailbreak
  • Harm/toxicity and unsafe content
  • PII and secrets leakage
  • Industry or policy-specific requirements

​
2) Map Coverage to Providers

  • Bud Sentinel for broad curated probe coverage and platform-native workflows.
  • Cloud provider moderation where external alignment or ecosystem constraints apply.

​
3) Build a Minimal Viable Probe Set

Start with high-impact probes only, then expand based on real incident and false-positive data.

​
4) Tune Thresholds with Data

Measure flag rates and business impact, then refine profile/probe thresholds.

​
Example Layered Policy

​
Rollout Pattern

  1. Pilot on one project.
  2. Monitor violations and false positives.
  3. Calibrate thresholds and probe selection.
  4. Replicate profile template across projects.

​
Governance Check Strategy

For agentic AI, extend provider and probe planning beyond input and output moderation. Map checks to the agent lifecycle: before the run starts, before or after model calls, before or after tool calls, and after the run completes. Add tool-access patterns for tools that should always be allowed, denied, escalated, or reviewed. Use filter conditions for narrow exceptions such as file type, environment, or data context.