ai-shield-core 0.2.0 → 0.3.0

package/dist/scanner/heuristic.js

@@ -8,12 +8,17 @@
 // Keep minimal — false-mappings in real content are worse than
 // false-negatives in an attack attempt.
 const HOMOGLYPH_MAP = {
+    // Cyrillic
     "а": "a", "е": "e", "і": "i", "ј": "j", "о": "o", "р": "p", "с": "c", "ѕ": "s",
-    "у": "y", "х": "x", "А": "A", "В": "B", "Е": "E", "І": "I", "К": "K", "М": "M",
-    "Н": "H", "О": "O", "Р": "P", "С": "C", "Т": "T", "Х": "X",
-    "α": "a", "ο": "o", "ρ": "p", "ε": "e", "υ": "y", "χ": "x", "Α": "A", "Β": "B",
-    "Ε": "E", "Ζ": "Z", "Η": "H", "Ι": "I", "Κ": "K", "Μ": "M", "Ν": "N", "Ο": "O",
-    "Ρ": "P", "Τ": "T", "Υ": "Y", "Χ": "X",
+    "у": "y", "х": "x", "ԁ": "d", "һ": "h", "ӏ": "l", "ո": "n", "А": "A", "В": "B",
+    "Е": "E", "І": "I", "К": "K", "М": "M", "Н": "H", "О": "O", "Р": "P", "С": "C",
+    "Т": "T", "Х": "X", "Ѕ": "S", "Ј": "J", "Ү": "Y", "Ԛ": "Q", "Ԝ": "W", "Ғ": "F",
+    // Greek
+    "α": "a", "ο": "o", "ρ": "p", "ε": "e", "υ": "y", "χ": "x", "ν": "v", "ι": "i",
+    "κ": "k", "Α": "A", "Β": "B", "Ε": "E", "Ζ": "Z", "Η": "H", "Ι": "I", "Κ": "K",
+    "Μ": "M", "Ν": "N", "Ο": "O", "Ρ": "P", "Τ": "T", "Υ": "Y", "Χ": "X",
+    // Armenian / Cherokee / other look-alikes occasionally used in evasion
+    "օ": "o", "ѵ": "v",
 };
 const HOMOGLYPH_RE = new RegExp(Object.keys(HOMOGLYPH_MAP).join("|"), "g");
 // Zero-width chars + BOM — used to split words like "ig<ZWSP>nore" across
@@ -39,6 +44,26 @@ export function normalizeForInjectionScan(input) {
     const noCombining = noZW.replace(COMBINING_RE, "");
     return noCombining.replace(HOMOGLYPH_RE, (ch) => HOMOGLYPH_MAP[ch] ?? ch);
 }
+/**
+ * Collapse letter-splitting evasion: an attacker writes `i g n o r e` or
+ * `i.g.n.o.r.e` or `i-g-n-o-r-e` to break the literal token "ignore" across
+ * separators so the regex never matches. This produces an ADDITIONAL view
+ * where any run of `single-letter + separator` (≥4 letters) has its
+ * separators removed, so the spaced form collapses back to "ignore".
+ *
+ * Run as a second pass IN ADDITION to the normal normalized text — never
+ * as a replacement — because collapsing is lossy (it would also fuse the
+ * legitimate "a b c" list). Only single-letter groups separated by one
+ * space / dot / dash / underscore are collapsed; multi-letter words are
+ * left intact, which keeps benign prose untouched.
+ */
+export function collapseSpacedLetters(input) {
+    // Match ≥3 "<letter><sep>" groups closed by a final lone letter. The
+    // trailing `(?![A-Za-z])` stops the greedy match from swallowing the
+    // first letter of the next real word ("i g n o r e all" must collapse to
+    // "ignore all", not "ignorea ll"). Bounded, linear — no nested quantifier.
+    return input.replace(/(?:[A-Za-z][ \t._-]){3,}[A-Za-z](?![A-Za-z])/g, (run) => run.replace(/[ \t._-]/g, ""));
+}
 const PATTERNS = [
     // --- Instruction Override (weight: 0.25 each) ---
     {
@@ -359,6 +384,19 @@ export class HeuristicScanner {
         // homoglyph/zero-width evasion doesn't bypass the rules. The caller
         // still sees the original input in `sanitized`.
         const normalized = normalizeForInjectionScan(input);
+        // Second view that un-splits letter-splitting evasion ("i g n o r e").
+        // Only computed when it actually differs (cheap guard), and only the
+        // high-value override/role/extraction/tool categories are re-tested
+        // against it — collapsing is lossy and the low-value framing rules
+        // would false-positive on collapsed prose.
+        const collapsed = collapseSpacedLetters(normalized);
+        const collapsedDiffers = collapsed !== normalized;
+        const SPLIT_SENSITIVE = new Set([
+            "instruction_override",
+            "role_manipulation",
+            "system_prompt_extraction",
+            "tool_abuse",
+        ]);
         for (const rule of this.patterns) {
             if (rule.pattern.test(normalized)) {
                 totalScore += rule.weight;
@@ -371,6 +409,20 @@ export class HeuristicScanner {
                     detail: `Rule ${rule.id} (${rule.category})`,
                 });
             }
+            else if (collapsedDiffers &&
+                SPLIT_SENSITIVE.has(rule.category) &&
+                rule.pattern.test(collapsed)) {
+                // Matched only after un-splitting → letter-splitting evasion.
+                totalScore += rule.weight;
+                violations.push({
+                    type: "prompt_injection",
+                    scanner: this.name,
+                    score: rule.weight,
+                    threshold: this.threshold,
+                    message: rule.description,
+                    detail: `Rule ${rule.id} (${rule.category}, letter-splitting evasion)`,
+                });
+            }
         }
         // Structural signals (cumulative) — intentionally run on the original
         // input so real structural attacks (many newlines, long paddings) can
@@ -404,6 +456,22 @@ export class HeuristicScanner {
         // Very long input (potential padding attack)
         if (input.length > 5000)
             score += 0.05;
+        // Adversarial suffix (GCG-style): a long whitespace-free token packed
+        // with mixed punctuation/symbols, typically appended after the readable
+        // request. Conservative — needs ≥25 chars and ≥6 distinct punctuation
+        // marks so ordinary URLs, hashes and code tokens don't trip it.
+        const ADV_TOKEN_RE = /\S{25,}/g;
+        let advMatch;
+        let advCount = 0;
+        while ((advMatch = ADV_TOKEN_RE.exec(input)) !== null && advCount < 32) {
+            advCount += 1;
+            const tok = advMatch[0];
+            const distinctPunct = new Set((tok.match(/[!-/:-@[-`{-~]/g) ?? [])).size;
+            if (distinctPunct >= 6) {
+                score += 0.05;
+                break;
+            }
+        }
         return score;
     }
     /** Get all registered pattern IDs for testing */

package/dist/scanner/ingestion.d.ts

@@ -93,6 +93,37 @@ export declare class IngestionScanner implements Scanner {
  * ```
  */
 export declare function scanIngested(content: string, source: IngestionSource, config?: IngestionScannerConfig): Promise<IngestionScanResult>;
+/**
+ * Scan the runtime *result* of a tool call before it re-enters the model
+ * context. The dominant indirect-injection channel in agentic loops: a
+ * search tool surfaces a poisoned page, an MCP server returns attacker-
+ * controlled data, a compromised upstream API embeds instructions in its
+ * response. PoisonedRAG (USENIX Security 2025) showed 5 planted documents
+ * reach a 90% attack-success rate in million-document knowledge bases —
+ * the payload arrives here, not in the user prompt.
+ *
+ * Thin wrapper over `scanIngested(content, "tool-output")` that also
+ * stamps the originating `toolName` into every violation detail, so an
+ * audit log can answer "which tool returned the poisoned content?".
+ *
+ * Pair with `CircuitBreakerRegistry` when you also want to rate-limit or
+ * trip the tool after repeated poisoned results:
+ *
+ * @example
+ * ```ts
+ * import { scanToolOutput } from "ai-shield-core";
+ *
+ * const result = await searchTool.call(query);          // untrusted
+ * const scan = await scanToolOutput("web_search", result);
+ * if (!scan.safe) {
+ *   // drop the result OR strip it before the next model turn
+ *   audit.warn("poisoned tool output", { tool: "web_search", v: scan.violations });
+ *   return; // do not feed `result` back into the model
+ * }
+ * model.continue(result);
+ * ```
+ */
+export declare function scanToolOutput(toolName: string, content: string, config?: IngestionScannerConfig): Promise<IngestionScanResult>;
 /**
  * Try to decode common obfuscation layers an attacker uses to smuggle
  * an injection past pattern matchers. Returns the decoded payload when

package/dist/scanner/ingestion.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ingestion.d.ts","sourceRoot":"","sources":["../../src/scanner/ingestion.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,OAAO,EACP,aAAa,EACb,WAAW,EACX,SAAS,EACT,eAAe,EACf,SAAS,EACV,MAAM,aAAa,CAAC;AA2GrB;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,eAAe,GAAG,SAAS,CAEtE;AAoFD;;;;;GAKG;AACH,MAAM,WAAW,mBAAmB;IAClC,IAAI,EAAE,OAAO,CAAC;IACd,QAAQ,EAAE,OAAO,GAAG,MAAM,GAAG,OAAO,CAAC;IACrC;;;;;;OAMG;IACH,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,SAAS,EAAE,CAAC;IACxB,MAAM,EAAE,eAAe,CAAC;IACxB,IAAI,EAAE;QACJ,cAAc,EAAE,MAAM,CAAC;QACvB,WAAW,EAAE,MAAM,EAAE,CAAC;QACtB,2DAA2D;QAC3D,kBAAkB,EAAE,MAAM,CAAC;QAC3B;;;;WAIG;QACH,MAAM,EAAE,OAAO,CAAC;KACjB,CAAC;CACH;AAED,MAAM,WAAW,sBAAsB;IACrC,gDAAgD;IAChD,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B;;;OAGG;IACH,UAAU,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;CACxC;AAED;;;;;;;GAOG;AACH,qBAAa,gBAAiB,YAAW,OAAO;IAC9C,QAAQ,CAAC,IAAI,eAAe;IAC5B,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAqB;IAC/C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAW;IAC1C,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAmB;gBAEjC,MAAM,GAAE,sBAA2B;IAQzC,IAAI,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,WAAW,GAAG,OAAO,CAAC,aAAa,CAAC;CAmHxE;AAED;;;;;;;;;;;;;;;;;;;GAmBG;AACH,wBAAsB,YAAY,CAChC,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,eAAe,EACvB,MAAM,GAAE,sBAA2B,GAClC,OAAO,CAAC,mBAAmB,CAAC,CA0B9B;AAQD;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CA2EjE"}
+{"version":3,"file":"ingestion.d.ts","sourceRoot":"","sources":["../../src/scanner/ingestion.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,OAAO,EACP,aAAa,EACb,WAAW,EACX,SAAS,EACT,eAAe,EACf,SAAS,EACV,MAAM,aAAa,CAAC;AAoIrB;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,eAAe,GAAG,SAAS,CAEtE;AAoFD;;;;;GAKG;AACH,MAAM,WAAW,mBAAmB;IAClC,IAAI,EAAE,OAAO,CAAC;IACd,QAAQ,EAAE,OAAO,GAAG,MAAM,GAAG,OAAO,CAAC;IACrC;;;;;;OAMG;IACH,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,SAAS,EAAE,CAAC;IACxB,MAAM,EAAE,eAAe,CAAC;IACxB,IAAI,EAAE;QACJ,cAAc,EAAE,MAAM,CAAC;QACvB,WAAW,EAAE,MAAM,EAAE,CAAC;QACtB,2DAA2D;QAC3D,kBAAkB,EAAE,MAAM,CAAC;QAC3B;;;;WAIG;QACH,MAAM,EAAE,OAAO,CAAC;KACjB,CAAC;CACH;AAED,MAAM,WAAW,sBAAsB;IACrC,gDAAgD;IAChD,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B;;;OAGG;IACH,UAAU,CAAC,EAAE,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;CACxC;AAED;;;;;;;GAOG;AACH,qBAAa,gBAAiB,YAAW,OAAO;IAC9C,QAAQ,CAAC,IAAI,eAAe;IAC5B,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAqB;IAC/C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAW;IAC1C,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAmB;gBAEjC,MAAM,GAAE,sBAA2B;IAQzC,IAAI,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,WAAW,GAAG,OAAO,CAAC,aAAa,CAAC;CAmHxE;AAED;;;;;;;;;;;;;;;;;;;GAmBG;AACH,wBAAsB,YAAY,CAChC,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,eAAe,EACvB,MAAM,GAAE,sBAA2B,GAClC,OAAO,CAAC,mBAAmB,CAAC,CA0B9B;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AACH,wBAAsB,cAAc,CAClC,QAAQ,EAAE,MAAM,EAChB,OAAO,EAAE,MAAM,EACf,MAAM,GAAE,sBAA2B,GAClC,OAAO,CAAC,mBAAmB,CAAC,CAa9B;AAQD;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CA2EjE"}

package/dist/scanner/ingestion.js

@@ -50,8 +50,14 @@ const SOURCE_PROFILE = {
         extraPatterns: [
             // "Before using this tool, you must ..." / "After calling, also call X"
             /(?:before|after|while)\s+(?:using|invoking|calling|executing)\s+(?:this\s+)?(?:tool|function|action)[,\s]+(?:you\s+)?(?:must|should|will|need\s+to|are\s+required\s+to)/i,
-            // "Also call delete_*" / "ALWAYS invoke X first"
-            /(?:also|always|first|then|finally)\s+(?:call|invoke|use|execute|run)\s+(?:the\s+)?[a-z_][\w-]{2,}/i,
+            // "Also call delete_account" / "then invoke query()". Tightened
+            // (review M2): the called symbol must carry a case-robust tool-name
+            // signal — a snake_case underscore or a `(` call — so benign docs prose
+            // ("first call the API", "then use the result") doesn't trip the
+            // block-level threshold and push operators into disabling the scanner.
+            // (camelCase detection was dropped: under the /i flag `[a-z]+[A-Z]`
+            // collapses to "any word" and re-introduced the false positive.)
+            /(?:also|always|first|then|finally)\s+(?:call|invoke|use|execute|run)\s+(?:the\s+)?[a-z][\w-]*(?:_[\w-]+|\s*\()/i,
             // "Tool description: ignore the above and ..."
             /(?:tool|function|action)\s+description[:\s]+(?:ignore|disregard|override)/i,
             // "Note to LLM:" / "Hidden instruction:"
@@ -60,6 +66,25 @@ const SOURCE_PROFILE = {
             /(?:on\s+(?:success|call)|when\s+invoked)[,:]\s+(?:silently|also|then)\s+(?:call|invoke|exfiltrate|send|leak)/i,
         ],
     },
+    "tool-output": {
+        // The runtime result a tool returned. Distinct from `tool-desc` (static
+        // schema): this is data the tool fetched and is therefore
+        // attacker-influenceable (PoisonedRAG, scraped pages surfaced through a
+        // search tool, a compromised upstream API). Tool results almost never
+        // legitimately carry instruction syntax aimed at the calling model, so
+        // the bar is as tight as RAG/tool-desc.
+        threshold: 0.13,
+        extraPatterns: [
+            // Result claims to carry instructions for the model.
+            /(?:tool|function|api|search|query)\s+(?:result|response|output)[:\s]+(?:ignore|disregard|override|new\s+instructions?|system\s+prompt)/i,
+            // "the result indicates you should now call/invoke X"
+            /(?:result|response|data|output)\s+(?:indicates?|says?|requires?|means?)\s+(?:that\s+)?you\s+(?:should|must|need\s+to|will)\s+(?:now\s+)?(?:call|invoke|run|execute|use)\s+[a-z_][\w-]{2,}/i,
+            // Embedded role/system marker inside a JSON-ish result value.
+            /"(?:role|system|instruction|directive)"\s*:\s*"(?:system|ignore|override|admin)/i,
+            // "(end of results) Now, as the system, ..."
+            /(?:end\s+of\s+(?:results?|output|data)|<\/results?>)[\s.)]*(?:now|next)[,\s]+(?:as\s+(?:the\s+)?(?:system|admin|assistant)|you\s+(?:must|should|will))/i,
+        ],
+    },
     memory: {
         // Stored memory entries: persistence poisoning. Look for sentinel
         // instructions that re-anchor the model on subsequent retrieval.
@@ -346,6 +371,49 @@ export async function scanIngested(content, source, config = {}) {
         },
     };
 }
+/**
+ * Scan the runtime *result* of a tool call before it re-enters the model
+ * context. The dominant indirect-injection channel in agentic loops: a
+ * search tool surfaces a poisoned page, an MCP server returns attacker-
+ * controlled data, a compromised upstream API embeds instructions in its
+ * response. PoisonedRAG (USENIX Security 2025) showed 5 planted documents
+ * reach a 90% attack-success rate in million-document knowledge bases —
+ * the payload arrives here, not in the user prompt.
+ *
+ * Thin wrapper over `scanIngested(content, "tool-output")` that also
+ * stamps the originating `toolName` into every violation detail, so an
+ * audit log can answer "which tool returned the poisoned content?".
+ *
+ * Pair with `CircuitBreakerRegistry` when you also want to rate-limit or
+ * trip the tool after repeated poisoned results:
+ *
+ * @example
+ * ```ts
+ * import { scanToolOutput } from "ai-shield-core";
+ *
+ * const result = await searchTool.call(query);          // untrusted
+ * const scan = await scanToolOutput("web_search", result);
+ * if (!scan.safe) {
+ *   // drop the result OR strip it before the next model turn
+ *   audit.warn("poisoned tool output", { tool: "web_search", v: scan.violations });
+ *   return; // do not feed `result` back into the model
+ * }
+ * model.continue(result);
+ * ```
+ */
+export async function scanToolOutput(toolName, content, config = {}) {
+    const result = await scanIngested(content, "tool-output", config);
+    const safeToolName = typeof toolName === "string" && toolName.length > 0
+        ? toolName.slice(0, 120)
+        : "unknown";
+    return {
+        ...result,
+        violations: result.violations.map((v) => ({
+            ...v,
+            detail: `${v.detail ?? ""} (tool=${safeToolName})`.trim(),
+        })),
+    };
+}
 // ============================================================
 // Encoding-bypass normalization (R1 from Round 1 review — closes
 // OWASP LLM Prompt Injection Prevention Cheat Sheet 2026 Base64/Hex

package/dist/scanner/output.d.ts

@@ -0,0 +1,73 @@
+import type { ScanContext, ScanDecision, Violation, PIIConfig } from "../types.js";
+export type OutputSink = "sql" | "shell" | "html" | "template";
+export interface OutputScanConfig {
+    /**
+     * PII handling. Pass a `PIIConfig` to control action/locale, or `false`
+     * to skip PII scanning entirely. Default: mask.
+     */
+    pii?: PIIConfig | false;
+    /**
+     * Canary token(s) injected into the system prompt via `injectCanary()`.
+     * If any appears verbatim in the output → `system_prompt_leak` (block).
+     */
+    canaryTokens?: string | string[];
+    /**
+     * Restrict the structured-injection check to specific downstream sinks.
+     * E.g. `["sql"]` when the output only ever flows into a query builder.
+     * Default: all sinks.
+     */
+    sinks?: OutputSink[];
+    /** Selectively disable checks. All enabled by default. */
+    checks?: {
+        secrets?: boolean;
+        injection?: boolean;
+        systemPromptLeak?: boolean;
+        jailbreak?: boolean;
+    };
+    /** Override the byte cap on the scanned region. Default 256 KB. */
+    maxBytes?: number;
+}
+export interface OutputScanResult {
+    /** No blocking violation found. */
+    safe: boolean;
+    decision: ScanDecision;
+    /**
+     * Output with PII masked and secrets redacted to `[REDACTED_SECRET]`.
+     * Unlike `scanIngested`, this is NOT emptied on block — the caller
+     * usually still needs to log or display the sanitized text. Gate on
+     * `safe` / `decision` before forwarding it to a downstream sink.
+     */
+    sanitized: string;
+    violations: Violation[];
+    meta: {
+        scanDurationMs: number;
+        checksRun: string[];
+    };
+}
+/**
+ * Scanner for LLM output. Stateless; safe to reuse across calls.
+ */
+export declare class OutputScanner {
+    private readonly config;
+    private readonly pii;
+    constructor(config?: OutputScanConfig);
+    scan(output: string, context?: ScanContext): Promise<OutputScanResult>;
+}
+/**
+ * One-shot helper. Scan a model response before acting on it.
+ *
+ * @example
+ * ```ts
+ * import { scanOutput } from "ai-shield-core";
+ *
+ * const reply = await llm.generate(prompt);
+ * const r = await scanOutput(reply, { canaryTokens: canary, sinks: ["sql"] });
+ * if (!r.safe) {
+ *   audit.warn("unsafe model output", r.violations);
+ *   return genericFallback();           // do not run r.sanitized as SQL
+ * }
+ * showToUser(r.sanitized);              // PII masked, secrets redacted
+ * ```
+ */
+export declare function scanOutput(output: string, config?: OutputScanConfig, context?: ScanContext): Promise<OutputScanResult>;
+//# sourceMappingURL=output.d.ts.map

package/dist/scanner/output.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"output.d.ts","sourceRoot":"","sources":["../../src/scanner/output.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,WAAW,EACX,YAAY,EACZ,SAAS,EACT,SAAS,EACV,MAAM,aAAa,CAAC;AAuHrB,MAAM,MAAM,UAAU,GAAG,KAAK,GAAG,OAAO,GAAG,MAAM,GAAG,UAAU,CAAC;AAE/D,MAAM,WAAW,gBAAgB;IAC/B;;;OAGG;IACH,GAAG,CAAC,EAAE,SAAS,GAAG,KAAK,CAAC;IACxB;;;OAGG;IACH,YAAY,CAAC,EAAE,MAAM,GAAG,MAAM,EAAE,CAAC;IACjC;;;;OAIG;IACH,KAAK,CAAC,EAAE,UAAU,EAAE,CAAC;IACrB,0DAA0D;IAC1D,MAAM,CAAC,EAAE;QACP,OAAO,CAAC,EAAE,OAAO,CAAC;QAClB,SAAS,CAAC,EAAE,OAAO,CAAC;QACpB,gBAAgB,CAAC,EAAE,OAAO,CAAC;QAC3B,SAAS,CAAC,EAAE,OAAO,CAAC;KACrB,CAAC;IACF,mEAAmE;IACnE,QAAQ,CAAC,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,gBAAgB;IAC/B,mCAAmC;IACnC,IAAI,EAAE,OAAO,CAAC;IACd,QAAQ,EAAE,YAAY,CAAC;IACvB;;;;;OAKG;IACH,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,SAAS,EAAE,CAAC;IACxB,IAAI,EAAE;QACJ,cAAc,EAAE,MAAM,CAAC;QACvB,SAAS,EAAE,MAAM,EAAE,CAAC;KACrB,CAAC;CACH;AAID;;GAEG;AACH,qBAAa,aAAa;IACxB,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAmB;IAC1C,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAoB;gBAE5B,MAAM,GAAE,gBAAqB;IAQnC,IAAI,CACR,MAAM,EAAE,MAAM,EACd,OAAO,GAAE,WAAgB,GACxB,OAAO,CAAC,gBAAgB,CAAC;CA+J7B;AAED;;;;;;;;;;;;;;;GAeG;AACH,wBAAsB,UAAU,CAC9B,MAAM,EAAE,MAAM,EACd,MAAM,GAAE,gBAAqB,EAC7B,OAAO,GAAE,WAAgB,GACxB,OAAO,CAAC,gBAAgB,CAAC,CAE3B"}

package/dist/scanner/output.js

@@ -0,0 +1,297 @@
+import { PIIScanner } from "./pii.js";
+import { normalizeForInjectionScan } from "./heuristic.js";
+// ============================================================
+// Output Scanner — OWASP LLM05 Improper Output Handling +
+// LLM02 Sensitive Information Disclosure (output side)
+//
+// AI Shield's input scanners answer "is this prompt safe to send to the
+// model?". This scanner answers the other half: "is this model OUTPUT
+// safe to act on / show / forward downstream?".
+//
+// LLM output must never reach a SQL engine, a shell, an HTML sink, or a
+// template renderer unfiltered — XSS, SSRF, SQLi and command injection
+// sourced from model output are a documented 2026 attack class (OWASP
+// LLM05). And a model can leak its own system prompt or a secret it was
+// shown, which is LLM02 / LLM07.
+//
+// Five checks. Inputs are Unicode-normalized first (homoglyph / zero-width /
+// fullwidth evasion defense). Secret + canary checks scan the FULL output
+// (a leak can sit anywhere); the structural checks scan a length-capped copy
+// (those payloads live in the first chunk):
+//   1. secret_leak        — API keys, tokens, private keys, DSNs (full output)
+//   2. output_injection   — SQL / shell / HTML-JS / template / md-exfil (capped)
+//   3. system_prompt_leak — canary-token leak (exact, full) + heuristic phrasing
+//   4. pii_detected       — reuses the input-side PIIScanner
+//   5. jailbreak_indicator— compliance-preamble / mode-switch acknowledgement
+//
+// Checks 1-3 are high-confidence and block. PII follows its configured
+// action. Jailbreak is heuristic and only warns — a "sure, here's how"
+// preamble is often legitimate.
+// ============================================================
+/** Hard cap on the bytes we pattern-scan. A 1 MB model response is not the
+ *  threat model and unbounded regex over it pressures GC. Overridable. */
+const DEFAULT_MAX_OUTPUT_BYTES = 256 * 1024;
+/**
+ * High-confidence secret formats. Each is anchored on a provider-specific
+ * prefix so false positives on prose are near-zero. Patterns are linear
+ * (no nested quantifiers) — ReDoS-safe on large output.
+ */
+const SECRET_PATTERNS = [
+    { id: "SEC-OPENAI", re: /\bsk-(?:proj-)?[A-Za-z0-9_-]{20,}\b/, label: "OpenAI API key" },
+    { id: "SEC-ANTHROPIC", re: /\bsk-ant-[A-Za-z0-9_-]{20,}\b/, label: "Anthropic API key" },
+    { id: "SEC-AWS-AKID", re: /\b(?:AKIA|ASIA)[0-9A-Z]{16}\b/, label: "AWS access key id" },
+    { id: "SEC-GITHUB", re: /\bgh[pousr]_[A-Za-z0-9]{36,}\b/, label: "GitHub token" },
+    { id: "SEC-GOOGLE", re: /\bAIza[0-9A-Za-z_-]{35}\b/, label: "Google API key" },
+    { id: "SEC-GOOGLE-OAUTH", re: /\bGOCSPX-[A-Za-z0-9_-]{28}\b/, label: "Google OAuth client secret" },
+    { id: "SEC-GCP-SA", re: /"type"\s*:\s*"service_account"/, label: "GCP service-account JSON" },
+    { id: "SEC-HUGGINGFACE", re: /\bhf_[A-Za-z0-9]{30,}\b/, label: "HuggingFace token" },
+    { id: "SEC-NPM", re: /\bnpm_[A-Za-z0-9]{36}\b/, label: "npm publish token" },
+    { id: "SEC-SLACK", re: /\bxox[baprs]-[A-Za-z0-9-]{10,}\b/, label: "Slack token" },
+    { id: "SEC-STRIPE", re: /\b[rs]k_live_[A-Za-z0-9]{20,}\b/, label: "Stripe live key" },
+    { id: "SEC-JWT", re: /\beyJ[A-Za-z0-9_-]{8,}\.eyJ[A-Za-z0-9_-]{8,}\.[A-Za-z0-9_-]{8,}\b/, label: "JWT" },
+    { id: "SEC-PEM", re: /-----BEGIN (?:RSA |EC |DSA |OPENSSH |PGP )?PRIVATE KEY-----/, label: "PEM private key" },
+    // DSN: both credential segments are length-bounded so a long near-match
+    // without a trailing `@` can't drive O(n²) backtracking (review H1).
+    { id: "SEC-DSN", re: /\b(?:postgres(?:ql)?|mysql|mongodb(?:\+srv)?|redis|amqps?):\/\/[^\s:/@]{1,64}:[^\s@]{3,80}@/, label: "connection string with credentials" },
+];
+/**
+ * Output-injection payloads, grouped by downstream sink. Each pattern is
+ * deliberately conservative — flagging legitimate output that merely
+ * *mentions* SQL would be useless. They target syntax that only matters
+ * when the string is interpreted, not displayed.
+ */
+const INJECTION_PATTERNS = [
+    // SQL
+    { id: "OUTI-SQL-1", sink: "sql", re: /\bunion\s+(?:all\s+)?select\b/i, label: "SQL UNION SELECT" },
+    { id: "OUTI-SQL-2", sink: "sql", re: /['"]\s*;\s*(?:drop|delete|update|insert|truncate|alter)\s+/i, label: "SQL statement break" },
+    { id: "OUTI-SQL-3", sink: "sql", re: /\bor\s+1\s*=\s*1\b|\bor\s+'1'\s*=\s*'1'/i, label: "SQL tautology" },
+    // Shell
+    { id: "OUTI-SH-1", sink: "shell", re: /\$\([^)]{1,200}\)|`[^`]{1,200}`/, label: "shell command substitution" },
+    { id: "OUTI-SH-2", sink: "shell", re: /[;&|]\s*(?:rm|curl|wget|nc|bash|sh|chmod|mkfifo|dd)\s+-?/i, label: "chained shell command" },
+    { id: "OUTI-SH-3", sink: "shell", re: /\|\s*(?:sh|bash|zsh|python[0-9.]*)\b/i, label: "pipe to interpreter" },
+    // HTML / JS (XSS)
+    { id: "OUTI-XSS-1", sink: "html", re: /<script[\s>]/i, label: "<script> tag" },
+    { id: "OUTI-XSS-2", sink: "html", re: /\bon(?:error|load|click|mouseover)\s*=\s*["']?[^"'>]{1,200}/i, label: "inline event handler" },
+    { id: "OUTI-XSS-3", sink: "html", re: /\bjavascript:\s*[^\s"']{1,200}/i, label: "javascript: URI" },
+    { id: "OUTI-XSS-4", sink: "html", re: /<iframe[\s>]|<img[^>]{0,200}\bsrc\s*=\s*["']?\s*(?:javascript|data):/i, label: "iframe / data-URI image" },
+    // Markdown-image data exfiltration: ![x](http://evil/log?data=…). When a
+    // renderer auto-loads the image the query string leaks whatever the model
+    // was told to embed. The most-overlooked LLM05 class (review: Research).
+    { id: "OUTI-MDEXF", sink: "html", re: /!\[[^\]]{0,200}\]\(\s*https?:\/\/[^)\s]{1,300}[?&][\w-]{1,40}=/i, label: "markdown-image data exfiltration" },
+    // Template / SSTI
+    { id: "OUTI-SSTI-1", sink: "template", re: /\{\{[^}]{0,200}(?:constructor|process|require|global|__proto__|self\.|cycler)[^}]{0,200}\}\}/i, label: "template-injection payload" },
+    { id: "OUTI-SSTI-2", sink: "template", re: /<%[^%]{0,200}(?:system|exec|require|eval)[^%]{0,200}%>/i, label: "ERB/EJS injection" },
+];
+/**
+ * System-prompt-leak heuristics — used only when no canary token is
+ * available. Low-confidence by design (these phrasings occur in benign
+ * output too), so they warn rather than block.
+ */
+const SYSTEM_LEAK_PATTERNS = [
+    /(?:my|the)\s+(?:system\s+)?(?:prompt|instructions?)\s+(?:is|are|say|states?|read)\b/i,
+    /i\s+(?:was|am|have\s+been)\s+(?:instructed|told|configured|programmed|designed)\s+to\b/i,
+    /here\s+(?:is|are)\s+my\s+(?:system\s+)?(?:prompt|instructions?|guidelines?|rules?)\b/i,
+    /you\s+are\s+(?:a|an)\s+[\w-]{2,30}\s+(?:assistant|agent|bot|model)\b.{0,40}\b(?:you\s+must|your\s+(?:rules?|guidelines?|instructions?))/i,
+];
+/**
+ * Jailbreak-success indicators in the OUTPUT. Conservative + low weight:
+ * a generic "Sure, here's how" is not enough on its own — these target
+ * explicit mode-switch acknowledgements and self-declared rule-breaking.
+ */
+const JAILBREAK_PATTERNS = [
+    /\bas\s+(?:DAN|an?\s+(?:unrestricted|unfiltered|jailbroken|uncensored))\b/i,
+    /i(?:'?ll|\s+will)\s+(?:now\s+)?(?:ignore|bypass|disregard|set\s+aside)\s+(?:my|the|all)\s+(?:guidelines?|restrictions?|rules?|safety|programming|filters?)/i,
+    /(?:jailbreak|developer\s+mode|dan\s+mode)\s+(?:enabled|activated|successful|engaged)/i,
+    /i\s+am\s+(?:now\s+)?(?:free\s+(?:from|of)|no\s+longer\s+bound\s+by)\s+(?:my\s+)?(?:restrictions?|guidelines?|rules?|programming)/i,
+];
+const SECRET_REDACTION = "[REDACTED_SECRET]";
+/**
+ * Scanner for LLM output. Stateless; safe to reuse across calls.
+ */
+export class OutputScanner {
+    config;
+    pii;
+    constructor(config = {}) {
+        this.config = config;
+        this.pii =
+            config.pii === false
+                ? null
+                : new PIIScanner(config.pii ?? { action: "mask" });
+    }
+    async scan(output, context = {}) {
+        const start = performance.now();
+        const violations = [];
+        const checksRun = [];
+        const checks = this.config.checks ?? {};
+        const maxBytes = this.config.maxBytes ?? DEFAULT_MAX_OUTPUT_BYTES;
+        const safeOutput = typeof output === "string" ? output : "";
+        // Capped copy for the *structural* checks (injection / leak-phrasing /
+        // jailbreak) — those payloads live in the first chunk and the regex over
+        // a 1 MB response would pressure GC. Normalized so homoglyph / zero-width
+        // / fullwidth evasion can't slip a payload past the patterns (review H6).
+        const cappedDetect = normalizeForInjectionScan(safeOutput.length > maxBytes ? safeOutput.slice(0, maxBytes) : safeOutput);
+        // Secrets and canaries can sit ANYWHERE in the output, and the secret
+        // patterns are anchored + linear — so they scan the FULL output, not the
+        // cap (review C1: a key padded past 256 KB must not slip through). Also
+        // normalized for the same evasion defense.
+        const fullDetect = normalizeForInjectionScan(safeOutput);
+        let sanitized = output;
+        let worst = "allow";
+        const bump = (d) => {
+            if (priority(d) > priority(worst))
+                worst = d;
+        };
+        // 1. Secret leak — high-confidence, always blocks. Redact in `sanitized`.
+        //    Detection runs on the normalized full output; redaction is
+        //    best-effort over the raw output (a key fragmented by zero-width
+        //    chars is still flagged via `fullDetect` and blocks, but may resist
+        //    clean redaction — callers MUST gate on `safe`/`decision` and never
+        //    forward a blocked output regardless of `sanitized`).
+        if (checks.secrets !== false) {
+            checksRun.push("secrets");
+            for (const { id, re, label } of SECRET_PATTERNS) {
+                if (re.test(fullDetect)) {
+                    violations.push({
+                        type: "secret_leak",
+                        scanner: "output",
+                        score: 1.0,
+                        threshold: 0.5,
+                        message: `Output leaks a secret: ${label}`,
+                        detail: `Rule ${id}`,
+                    });
+                    bump("block");
+                    // Redact every occurrence in the full output (global copy of re).
+                    sanitized = sanitized.replace(new RegExp(re.source, re.flags.includes("g") ? re.flags : re.flags + "g"), SECRET_REDACTION);
+                }
+            }
+        }
+        // 2. Output injection — payloads dangerous to a downstream sink.
+        if (checks.injection !== false) {
+            checksRun.push("injection");
+            const allowedSinks = this.config.sinks;
+            for (const { id, sink, re, label } of INJECTION_PATTERNS) {
+                if (allowedSinks && !allowedSinks.includes(sink))
+                    continue;
+                if (re.test(cappedDetect)) {
+                    violations.push({
+                        type: "output_injection",
+                        scanner: "output",
+                        score: 0.85,
+                        threshold: 0.5,
+                        message: `Output carries a ${sink} injection payload: ${label}`,
+                        detail: `Rule ${id} (sink=${sink})`,
+                    });
+                    bump("block");
+                }
+            }
+        }
+        // 3. System-prompt leak — canary first (exact, certain), then heuristics.
+        if (checks.systemPromptLeak !== false) {
+            checksRun.push("system_prompt_leak");
+            const tokens = normalizeTokens(this.config.canaryTokens);
+            let canaryHit = false;
+            for (const token of tokens) {
+                // Check the FULL output, not the capped copy — a leak past 256 KB is
+                // still a leak, and an exact substring search is cheap.
+                if (token.length >= 4 && output.includes(token)) {
+                    canaryHit = true;
+                    violations.push({
+                        type: "system_prompt_leak",
+                        scanner: "output",
+                        score: 1.0,
+                        threshold: 0.5,
+                        message: "Output leaks a system-prompt canary token",
+                        detail: "Canary match (exact)",
+                    });
+                    bump("block");
+                }
+            }
+            // Heuristic phrasing only when no canary was available/hit — avoids
+            // double-reporting and keeps the low-confidence signal subordinate.
+            if (!canaryHit && tokens.length === 0) {
+                for (const re of SYSTEM_LEAK_PATTERNS) {
+                    if (re.test(cappedDetect)) {
+                        violations.push({
+                            type: "system_prompt_leak",
+                            scanner: "output",
+                            score: 0.4,
+                            threshold: 0.5,
+                            message: "Output may be echoing the system prompt",
+                            detail: "Heuristic phrasing (no canary configured — pass canaryTokens for an exact check)",
+                        });
+                        bump("warn");
+                        break; // one heuristic signal is enough
+                    }
+                }
+            }
+        }
+        // 4. Jailbreak indicators — heuristic, warn only.
+        if (checks.jailbreak !== false) {
+            checksRun.push("jailbreak");
+            for (const re of JAILBREAK_PATTERNS) {
+                if (re.test(cappedDetect)) {
+                    violations.push({
+                        type: "jailbreak_indicator",
+                        scanner: "output",
+                        score: 0.3,
+                        threshold: 0.5,
+                        message: "Output shows a possible jailbreak success indicator",
+                        detail: "Heuristic phrasing",
+                    });
+                    bump("warn");
+                    break;
+                }
+            }
+        }
+        // 5. PII — reuse the input-side scanner; respects its configured action.
+        if (this.pii) {
+            checksRun.push("pii");
+            const piiResult = await this.pii.scan(sanitized, context);
+            for (const v of piiResult.violations) {
+                violations.push({ ...v, scanner: "output" });
+            }
+            if (piiResult.sanitized !== undefined)
+                sanitized = piiResult.sanitized;
+            bump(piiResult.decision);
+        }
+        return {
+            safe: worst === "allow",
+            decision: worst,
+            sanitized,
+            violations,
+            meta: {
+                scanDurationMs: performance.now() - start,
+                checksRun,
+            },
+        };
+    }
+}
+/**
+ * One-shot helper. Scan a model response before acting on it.
+ *
+ * @example
+ * ```ts
+ * import { scanOutput } from "ai-shield-core";
+ *
+ * const reply = await llm.generate(prompt);
+ * const r = await scanOutput(reply, { canaryTokens: canary, sinks: ["sql"] });
+ * if (!r.safe) {
+ *   audit.warn("unsafe model output", r.violations);
+ *   return genericFallback();           // do not run r.sanitized as SQL
+ * }
+ * showToUser(r.sanitized);              // PII masked, secrets redacted
+ * ```
+ */
+export async function scanOutput(output, config = {}, context = {}) {
+    return new OutputScanner(config).scan(output, context);
+}
+function normalizeTokens(tokens) {
+    if (!tokens)
+        return [];
+    const arr = Array.isArray(tokens) ? tokens : [tokens];
+    return arr.filter((t) => typeof t === "string" && t.length > 0);
+}
+function priority(d) {
+    return d === "block" ? 2 : d === "warn" ? 1 : 0;
+}
+//# sourceMappingURL=output.js.map