Guardrails

Built-in, language-agnostic guardrails — PII, prompt-injection, and wrong-language — that register on an agent like any hand-written guardrail.

Factories

• augments.adk.guardrails.pii_guardrail
• augments.adk.guardrails.injection_scan_guardrail
• augments.adk.guardrails.semantic_scan_guardrail
• augments.adk.guardrails.wrong_language_guardrail

Scanners

• augments.adk.guardrails.PatternScanner
• augments.adk.guardrails.SemanticScanner
• augments.adk.guardrails.SemanticMatch

Helpers

• augments.adk.guardrails.mask_pii_spans
• augments.adk.guardrails.fence_untrusted_text
• augments.adk.guardrails.detect_wrong_language

Defaults

• augments.adk.guardrails.DEFAULT_PII_MASK
• augments.adk.guardrails.DEFAULT_INJECTION_EXEMPLARS

Three further defaults, spelled out because their content is the interesting part:

• DEFAULT_PII_PATTERNS — cheap, deterministic regex markers for the common injected identifiers: email addresses (ASCII and internationalized), URLs, and phone numbers. Override via the patterns argument of pii_guardrail.
• DEFAULT_INJECTION_PATTERNS — high-confidence injection markers: broad English phrases plus the classic "ignore the previous instructions" signature across FR/DE/ES/PT/RU/ZH/JA/HI/AR. Override via the patterns argument of injection_scan_guardrail.
• DEFAULT_LANGUAGE_CODES — target-language name to ISO 639-1 code, spanning every language the detector supports. Override via the language_codes argument of wrong_language_guardrail.

Agent-level guardrail configuration is documented under the Guardrails guide.