Powered by

PII/PHI Guardrails[NEW]

Overview

PII/PHI Guardrails apply runtime privacy enforcement at the agent level, with regulatory presets, custom entity definitions, and bi-directional detection across inputs and outputs. They help keep raw PII/PHI out of agent outputs, traces, logs, and API responses.

PII/PHI Guardrails at the agent level operate independently of Knowledge Garden rules. To protect data before it is embedded in a Knowledge Garden, see EKG PII/PHI Guardrails.

Key Features

  • Detection methods: Regex, Deep Learning, and LLM-based, with configurable combinations
  • Regulatory presets: One-click presets for HIPAA (18 identifiers) and GDPR
  • Preconfigured entities: Common PII entities with default detection and masking strategies
  • Custom entities: Custom entity definitions with regex-based detection and masking or blocking behavior
  • Runtime enforcement: Applied before prompt construction, response rendering, logging, and tool calls
  • Masking conflict resolution: Automatic priority hierarchy (Tokenization > Partial > Full)
  • Coverage: Applicable to agent inputs, including on-demand uploaded files

Core Concepts

Guardrail Execution Order

Guardrails run in a structured order to minimize added latency:

  1. PII-PHI Masking executes first, sequentially.
  2. Injection Defense and Jailbreak Detection then run in parallel.

Sub-Agent Independence

Sub-agents do not inherit the guardrail configuration of their parent agent. Each sub-agent operates at its own configured guardrail level, which prevents unintended data exposure.

Role-Based Visibility

What each user sees depends on where they look and their role:

Where Who What they see
Chat Window and Test & Debug Agent Builders and users Original, unmasked inputs
Guardrail execution steps in Traces Agent Builders and Admins Strictly masked content
Governance Admins All user inputs and LLM responses in masked state
Governance (Chat History, Traces, Tool Outputs) Users with the Guardrail Enabled User policy Unmasked user queries; LLM responses stay masked

Guardrail Privileged User Policy

The Guardrail Enabled User policy lets designated compliance users view unmasked user queries in Governance, for incident investigation and audit purposes.

  • Grants an unmasked view of user queries in Chat History and Traces by default
  • LLM responses continue to display in a masked state, as per the guardrails configuration
  • Non-privileged users (Admins, Agent Builders, GenAI Users, Consumers) continue to see masked queries
  • The privilege is read-only: it gives no access to modify guardrail configurations or agent settings
  • Applies across Chat History, Traces, and Tool Outputs in Governance

Refer to User Policies to know more.

Benefits

  • Supports HIPAA and GDPR requirements through ready-made presets, reducing the need for custom engineering
  • Provides flexible, layered privacy controls that operate independently of Knowledge Garden rules
  • Reduces compliance risk by applying sanitization before prompt construction
  • Ensures a consistent compliance posture across all governance views, regardless of user role