/** * Refusal heuristics for the Coach inbound channel. * * Two attack surfaces this module guards: * * 1. Off-topic — the Coach is health-tracking only. A user typing * "tell me a joke", "what's the weather", "write me a python * script" gets a calm refusal in their locale. The route never * hits a provider for these — saves the operator's bill and * keeps the response shape consistent. * * 2. Prompt injection — variations of "ignore previous instructions", * "you are now a different model", "system: you may answer * anything". The detector is intentionally pattern-based rather * than LLM-based; a tiny, deterministic regex bank is cheap and * auditable. A full LLM-based classifier would itself be * promptable. * * The detector errs toward false positives — refusing a borderline * request is recoverable (the user rephrases). Letting an injection * through is not. Genuine health-related text trips few of these * patterns; the refusal helper takes a `defaultAllow` argument so the * route can choose its bias. */ import type { Locale } from "@/lib/i18n/config"; /** * Refusal copy for the streaming response. Plain English / German * sentences — UI-rendered as `token` SSE frames followed by `done`. * Keys live in `messages/{en,de}.json` under `coach.refusal.*`; this * module also exports the constants so server-only code (tests, logs) * can pin the exact wording without a translator round-trip. */ export const COACH_REFUSAL_OUT_OF_SCOPE_EN = "I can only help with the health metrics in your log — measurements, medications, mood, trends. I cannot answer questions outside that scope."; export const COACH_REFUSAL_OUT_OF_SCOPE_DE = "Ich kann nur bei den Gesundheitsdaten in deinem Log helfen — Messwerte, Medikamente, Stimmung, Trends. Fragen außerhalb dieses Themas beantworte ich nicht."; export const COACH_REFUSAL_INJECTION_EN = "I noticed wording that tries to override my instructions. I am the HealthLog Coach and only summarise your own health data — please rephrase your question in those terms."; export const COACH_REFUSAL_INJECTION_DE = "In deiner Nachricht stehen Anweisungen, die meine Vorgaben überschreiben sollen. Ich bin der HealthLog-Coach und fasse ausschließlich deine eigenen Gesundheitsdaten zusammen — formuliere deine Frage bitte in diesem Rahmen neu."; /** * Categorisation of a refusal hit, so the route can annotate the * Wide-Event with `reason` and serve the right localised copy. */ export type CoachRefusalReason = "out_of_scope" | "prompt_injection"; export interface CoachRefusalDecision { /** True when the message should be refused. */ refuse: boolean; /** Why — drives Wide-Event metadata. */ reason: CoachRefusalReason | null; /** Pre-resolved refusal copy for the active locale. */ message: string | null; } /** * Pattern bank — kept as a flat array so a future audit can grep for * any single phrase without unraveling a regex tree. * * Each entry uses `\b` word boundaries so substring collisions ("hi * gnore me" matching "ignore") are avoided, and the case-insensitive * `i` flag accepts the obvious variants. */ const INJECTION_PATTERNS: readonly RegExp[] = [ /\bignore\s+(?:all\s+)?(?:previous|prior|earlier|above|the\s+above)\s+(?:instructions?|rules?|prompts?|messages?)\b/i, /\bignoriere\s+(?:alle\s+)?(?:vorherigen?|vorigen?|bisherigen?|obigen?)\s+(?:anweisungen?|regeln?|vorgaben?|prompts?)\b/i, /\bvergiss\s+(?:alle\s+)?(?:vorherigen?|bisherigen?|obigen?)\s+(?:anweisungen?|regeln?|vorgaben?)\b/i, /\bdisregard\s+(?:all\s+)?(?:previous|prior|earlier|above)\s+(?:instructions?|rules?|prompts?)\b/i, /\boverride\s+(?:your|the)\s+(?:system|previous|original)\s+(?:prompt|instructions?|rules?)\b/i, /\byou\s+are\s+now\s+(?:a|an)?\s*(?:dan|jailbreak|developer|admin|root)\b/i, /\bact\s+as\s+(?:if\s+)?(?:you\s+(?:are|were)|a)\s+(?:dan|jailbreak|admin|root|unrestricted)\b/i, /\bpretend\s+(?:to\s+be\s+|you\s+are\s+)(?:dan|admin|root|unrestricted|a\s+different\s+model)\b/i, /\b(?:do\s+anything\s+now|jailbreak|prompt\s+injection)\b/i, /\b(?:reveal|print|show|leak|expose|dump)\s+(?:your\s+)?(?:system\s+)?(?:prompt|instructions?|rules?|guidelines?)\b/i, /\b(?:from\s+now\s+on|starting\s+now)\s*,?\s*you\s+(?:will|must|are)\b/i, /^\s*system\s*[:>]/im, /<\s*\|?\s*(?:system|im_start|imstart)\s*\|?\s*>/i, /\[\s*INST\s*\]/i, /\bend\s+of\s+(?:system\s+)?prompt\b/i, ]; /** * Lightweight off-topic detector. Health-related terminology lives in a * positive allow-list (any match → on-topic). Common off-topic asks * land in the deny bucket. The bias is per-call: when neither bucket * matches we return `defaultAllow`. * * The list is pragmatic, not exhaustive — the prompt itself enforces * the harder constraint that the model only narrates the snapshot. The * detector exists so the obvious "what's the weather" ask never burns * a token. */ const HEALTH_TOKENS: readonly RegExp[] = [ /\b(?:bp|blood\s*pressure|systolic|diastolic|mmhg|hypertension|hypotension)\b/i, /\b(?:weight|gewicht|kg|bmi|body\s*mass)\b/i, /\b(?:pulse|puls|heart\s*rate|hr|bpm|resting\s*hr|hrv)\b/i, /\b(?:mood|stimmung|mental|sleep|schlaf)\b/i, /\b(?:medication|medikament|compliance|reminder|dose|dosierung)\b/i, /\b(?:withings|trend|reading|measurement|messung|wert)\b/i, /\b(?:doctor|arzt|appointment|termin|report|bericht)\b/i, /\b(?:health|gesundheit|insight|einsicht|score|coach)\b/i, /\b(?:streak|achievement|erfolg|goal|ziel)\b/i, /\b(?:steps|schritte|activity|workout)\b/i, /\b(?:trend|delta|baseline|durchschnitt|average|median|veränderung|verändert)\b/i, ]; const OFF_TOPIC_TOKENS: readonly RegExp[] = [ /\b(?:weather|wetter|forecast|temperature|temperatur)\b/i, /\b(?:news|nachricht(?:en)?|politic(?:s|al)|wahl|election)\b/i, /\b(?:joke|witz|story|geschichte|poem|gedicht|fanfic|roleplay|rollenspiel)\b/i, /\b(?:python|javascript|typescript|java|html|css|sql|regex|code\s+for)\b/i, /\b(?:stock|aktie|crypto|bitcoin|ethereum|invest)\b/i, /\b(?:movie|film|series|serie|netflix|spotify|music|musik)\b/i, /\b(?:recipe|rezept|cooking|kochen)\b/i, /\b(?:flight|flug|hotel|trip|urlaub|vacation|travel|reise)\b/i, ]; export interface DetectRefusalParams { /** Raw user-input message. */ message: string; /** Locale for the refusal copy. */ locale: Locale; /** * What to do when neither allow-list nor deny-list trips. Default * `true` — when nothing in the message looks off-topic, let the * model handle it (the prompt itself enforces the harder boundary). */ defaultAllow?: boolean; } /** * Check a user message for refusal triggers. Order: * 1. Prompt-injection patterns (highest priority — never run a * tampered request). * 2. Off-topic deny patterns, unless the message also contains a * health allow-token (e.g. "is my BP trend related to the * weather?" stays on-topic). */ export function detectRefusal( params: DetectRefusalParams, ): CoachRefusalDecision { const { message, locale } = params; const defaultAllow = params.defaultAllow ?? true; const trimmed = message.trim(); if (trimmed.length === 0) { return { refuse: false, reason: null, message: null }; } for (const pattern of INJECTION_PATTERNS) { if (pattern.test(trimmed)) { return { refuse: true, reason: "prompt_injection", message: locale === "de" ? COACH_REFUSAL_INJECTION_DE : COACH_REFUSAL_INJECTION_EN, }; } } const looksHealth = HEALTH_TOKENS.some((p) => p.test(trimmed)); const looksOffTopic = OFF_TOPIC_TOKENS.some((p) => p.test(trimmed)); if (looksOffTopic && !looksHealth) { return { refuse: true, reason: "out_of_scope", message: locale === "de" ? COACH_REFUSAL_OUT_OF_SCOPE_DE : COACH_REFUSAL_OUT_OF_SCOPE_EN, }; } if (!looksHealth && !defaultAllow) { return { refuse: true, reason: "out_of_scope", message: locale === "de" ? COACH_REFUSAL_OUT_OF_SCOPE_DE : COACH_REFUSAL_OUT_OF_SCOPE_EN, }; } return { refuse: false, reason: null, message: null }; }