@vpdeva/blackwall-llm-shield-js 0.2.4 → 0.6.0

package/src/scorecard.js

@@ -1,14 +1,61 @@
 #!/usr/bin/env node
 
+const fs = require('fs');
+const path = require('path');
 const { BlackwallShield, runRedTeamSuite } = require('./index');
 
+function parseScalar(value = '') {
+  const trimmed = String(value).trim();
+  if (trimmed === 'true') return true;
+  if (trimmed === 'false') return false;
+  if (/^-?\d+$/.test(trimmed)) return Number(trimmed);
+  return trimmed.replace(/^['"]|['"]$/g, '');
+}
+
+function loadConfig(configPath) {
+  const raw = fs.readFileSync(configPath, 'utf8');
+  if (configPath.endsWith('.json')) return JSON.parse(raw);
+  const result = {};
+  let currentList = null;
+  let currentItem = null;
+  raw.split(/\r?\n/).forEach((line) => {
+    if (!line.trim() || line.trim().startsWith('#')) return;
+    if (/^[A-Za-z]/.test(line) && line.includes(':')) {
+      const [key, rest] = line.split(/:\s*/, 2);
+      if (!rest) {
+        result[key.trim()] = [];
+        currentList = result[key.trim()];
+        currentItem = null;
+      } else {
+        result[key.trim()] = parseScalar(rest);
+      }
+      return;
+    }
+    const itemMatch = line.match(/^\s*-\s+([A-Za-z0-9_]+):\s*(.+)$/);
+    if (itemMatch && currentList) {
+      currentItem = { [itemMatch[1]]: parseScalar(itemMatch[2]) };
+      currentList.push(currentItem);
+      return;
+    }
+    const propMatch = line.match(/^\s+([A-Za-z0-9_]+):\s*(.+)$/);
+    if (propMatch && currentItem) currentItem[propMatch[1]] = parseScalar(propMatch[2]);
+  });
+  return result;
+}
+
 async function main() {
+  const args = process.argv.slice(2);
+  const validateMode = args[0] === 'validate';
+  const configIndex = args.indexOf('--config');
+  const configPath = configIndex >= 0 ? path.resolve(args[configIndex + 1]) : null;
+  const config = configPath ? loadConfig(configPath) : {};
   const shield = new BlackwallShield({
     blockOnPromptInjection: true,
     promptInjectionThreshold: process.env.BLACKWALL_PROMPT_THRESHOLD || 'high',
     shadowMode: process.env.BLACKWALL_SHADOW_MODE === 'true',
+    ...config,
   });
-  const scorecard = await runRedTeamSuite({ shield, metadata: { source: 'cli' } });
+  const scorecard = await runRedTeamSuite({ shield, metadata: { source: 'cli', mode: validateMode ? 'validate' : 'run' } });
   process.stdout.write(`${JSON.stringify(scorecard, null, 2)}\n`);
 }
 

package/src/semantic.js

@@ -4,6 +4,13 @@ class TransformersIntentScorer {
   constructor(classifier, options = {}) {
     this.classifier = classifier;
     this.threshold = options.threshold || 0.5;
+    this.labelMap = options.labelMap || {
+      jailbreak: 'jailbreak',
+      prompt_injection: 'prompt_injection',
+      unsafe: 'unsafe',
+      injection: 'prompt_injection',
+      malicious: 'unsafe',
+    };
   }
 
   async score(text) {
@@ -12,7 +19,8 @@ class TransformersIntentScorer {
     const matches = [];
     let total = 0;
     for (const item of items) {
-      const label = String(item.label || '').toLowerCase();
+      const rawLabel = String(item.label || '').toLowerCase();
+      const label = this.labelMap[rawLabel] || rawLabel;
       const score = Number(item.score || 0);
       if (['jailbreak', 'prompt_injection', 'unsafe'].includes(label) && score >= this.threshold) {
         const weighted = Math.min(40, Math.round(score * 40));
@@ -32,7 +40,7 @@ class TransformersIntentScorer {
 async function createOptionalLocalIntentScorer(options = {}) {
   try {
     const { pipeline } = require('@xenova/transformers');
-    const classifier = await pipeline('text-classification', options.model || 'Xenova/distilbert-base-uncased-finetuned-sst-2-english');
+    const classifier = await pipeline('text-classification', options.model || 'ProtectAI/deberta-v3-base-prompt-injection-v2');
     return new TransformersIntentScorer(classifier, options);
   } catch {
     return new LightweightIntentScorer(options);