Changelog

All notable changes to @edwinfom/ai-guard are documented here.

v0.2.1 — 2026-04-13

registerModelPricing() — Register custom model pricing at runtime. Any model name, any price per 1M tokens.
New models in KnownModel — gemini-2.5-flash, gemini-2.5-pro, gpt-4.1, gpt-4.1-mini, claude-3-7-sonnet-20250219 added to the built-in pricing table.
extractJSON() and repairJSON() — New standalone exports from @edwinfom/ai-guard/schema for lower-level JSON repair without schema validation.
estimateTokensByChars() — New export from @edwinfom/ai-guard/budget for character-based token estimation.
getPatternsForTargets() — New export from @edwinfom/ai-guard/pii to retrieve the raw regex patterns for specific PII types.
InjectionResult type — Now exported from @edwinfom/ai-guard/injection.
RedactionResult type — Now exported from @edwinfom/ai-guard/pii.
Interactive Playground — Test all standalone modules directly in the documentation without any API key.

registerModelPricing now persists across CJS/ESM module boundaries via globalThis.
buildUsage correctly falls back to character-based estimation when real token counts are unavailable.

Canary Tokens — Invisible markers detect system prompt leakage at runtime
Content Policy — Detects toxicity, hate speech, violence, self-harm, sexual content
Hallucination Detection — Named-entity grounding check for RAG pipelines
Rate Limiter — Per-user sliding-window request and token limits
Audit Log — Structured callback with every protect() call
Added French PII types: NIR, SIRET, SIREN, passport, date of birth

Schema Level 2 now uses jsonrepair instead of custom regex extractor (100+ more patterns covered)