muaddib-scanner 2.10.31 → 2.10.33

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "muaddib-scanner",
-  "version": "2.10.31",
+  "version": "2.10.33",
   "description": "Supply-chain threat detection & response for npm & PyPI/Python",
   "main": "src/index.js",
   "bin": {
@@ -55,7 +55,7 @@
   },
   "devDependencies": {
     "@eslint/js": "10.0.1",
-    "eslint": "10.0.3",
+    "eslint": "10.1.0",
     "eslint-plugin-security": "^4.0.0",
     "globals": "17.4.0"
   }

package/src/ml/classifier.js

@@ -21,6 +21,10 @@ const { extractFeatures } = require('./feature-extractor.js');
 // Lazy-loaded models (allows resetModel for testing)
 let _model = undefined; // undefined = not yet loaded, null = absent
 let _bundlerModel = undefined; // undefined = not yet loaded, null = absent
+let _shadowModel = undefined; // undefined = not yet loaded, null = absent
+
+// Shadow mode stats (reset on model reload)
+const _shadowStats = { total: 0, agree: 0, disagree: 0 };
 
 // High-confidence malice types that must NEVER be suppressed by ML
 const HC_TYPES = new Set([
@@ -94,6 +98,80 @@ function resetBundlerModel() {
   _bundlerModel = undefined;
 }
 
+// --- Shadow model (ML1 v2, logs only, no filtering) ---
+
+/**
+ * Load shadow model from model-trees-shadow.js. Returns model object or null.
+ * Shadow model runs in parallel with the main model for comparison.
+ */
+function loadShadowModel() {
+  if (_shadowModel !== undefined) return _shadowModel;
+  try {
+    _shadowModel = require('./model-trees-shadow.js') || null;
+  } catch {
+    _shadowModel = null;
+  }
+  return _shadowModel;
+}
+
+function isShadowModelAvailable() {
+  return loadShadowModel() !== null;
+}
+
+function resetShadowModel() {
+  _shadowModel = undefined;
+  _shadowStats.total = 0;
+  _shadowStats.agree = 0;
+  _shadowStats.disagree = 0;
+}
+
+/**
+ * Run shadow model prediction and log comparison with main model.
+ * Never affects the actual classification decision.
+ *
+ * @param {Object} result - scan result
+ * @param {Object} meta - enriched metadata
+ * @param {string} mainPrediction - the main model's prediction
+ * @param {number} mainProbability - the main model's probability
+ * @param {string} packageName - for logging
+ */
+function runShadowComparison(result, meta, mainPrediction, mainProbability, packageName) {
+  const shadow = loadShadowModel();
+  if (!shadow) return;
+
+  const features = extractFeatures(result, meta || {});
+  const values = new Array(shadow.features.length);
+  for (let i = 0; i < shadow.features.length; i++) {
+    values[i] = features[shadow.features[i]] || 0;
+  }
+
+  let margin = 0;
+  for (const tree of shadow.trees) {
+    margin += traverseTree(tree, values);
+  }
+
+  const shadowProb = sigmoid(margin);
+  const shadowPred = shadowProb >= shadow.threshold ? 'malicious' : 'clean';
+
+  _shadowStats.total++;
+  if (shadowPred === mainPrediction) {
+    _shadowStats.agree++;
+  } else {
+    _shadowStats.disagree++;
+    console.log(`[ML-SHADOW] Disagreement on ${packageName}: main=${mainPrediction}(${mainProbability}) shadow=${shadowPred}(${Math.round(shadowProb * 1000) / 1000}) [${_shadowStats.disagree}/${_shadowStats.total} disagree]`);
+  }
+
+  // Periodic summary every 100 classifications
+  if (_shadowStats.total % 100 === 0) {
+    const agreeRate = ((_shadowStats.agree / _shadowStats.total) * 100).toFixed(1);
+    console.log(`[ML-SHADOW] Stats: ${_shadowStats.total} total, ${agreeRate}% agree, ${_shadowStats.disagree} disagree`);
+  }
+}
+
+function getShadowStats() {
+  return { ..._shadowStats };
+}
+
 /**
  * Sigmoid function: maps raw margin to probability [0, 1].
  * @param {number} x - raw margin (sum of tree outputs)
@@ -276,9 +354,18 @@ function classifyPackage(result, meta) {
   const featureValues = buildFeatureVector(result, meta);
   const { probability, prediction } = predict(featureValues);
 
+  const roundedProb = Math.round(probability * 1000) / 1000;
+
+  // Shadow model comparison (log-only, never affects decision)
+  if (isShadowModelAvailable()) {
+    const pkgName = (result && result.summary && result.summary.packageName) ||
+                    (meta && meta.name) || 'unknown';
+    runShadowComparison(result, meta, prediction, roundedProb, pkgName);
+  }
+
   return {
     prediction,
-    probability: Math.round(probability * 1000) / 1000,
+    probability: roundedProb,
     reason: prediction === 'clean' ? 'ml_clean' : 'ml_malicious'
   };
 }
@@ -298,5 +385,10 @@ module.exports = {
   resetBundlerModel,
   loadBundlerModel,
   predictBundler,
-  buildBundlerFeatureVector
+  buildBundlerFeatureVector,
+  // Shadow model (ML1 v2, log-only comparison)
+  isShadowModelAvailable,
+  resetShadowModel,
+  loadShadowModel,
+  getShadowStats
 };

package/src/ml/feature-extractor.js

@@ -13,7 +13,11 @@
 
 // Top threat types by frequency in production (covers ~95% of all findings).
 // Types not in this list are aggregated into `threat_type_other`.
+// v2.10.32: expanded from 31 to 47 types — code exec bypasses, IoC, GlassWorm,
+// obfuscation patterns, module graph sinks. New features will be 0 for pre-existing
+// JSONL records; SHAP feature selection handles sparsity gracefully.
 const TOP_THREAT_TYPES = [
+  // --- Original 31 types ---
   'suspicious_dataflow',
   'env_access',
   'sensitive_string',
@@ -44,7 +48,29 @@ const TOP_THREAT_TYPES = [
   'curl_exec',
   'reverse_shell',
   'binary_dropper',
-  'mcp_config_injection'
+  'mcp_config_injection',
+  // --- Code execution bypasses (v2.9.x–v2.10.x) ---
+  'vm_code_execution',
+  'vm_dynamic_code',
+  'dangerous_constructor',
+  'module_load_bypass',
+  'require_process_mainmodule',
+  'proxy_globalthis_intercept',
+  'reflect_bind_code_execution',
+  // --- IoC / supply chain ---
+  'known_malicious_package',
+  'known_malicious_hash',
+  // --- GlassWorm (Unicode + Blockchain C2) ---
+  'unicode_invisible_injection',
+  'blockchain_c2_resolution',
+  // --- Shell / exec patterns ---
+  'dangerous_exec',
+  'node_inline_exec',
+  // --- Obfuscation patterns ---
+  'js_obfuscation_pattern',
+  // --- Module graph / WASM ---
+  'suspicious_module_sink',
+  'wasm_host_sink'
 ];
 
 const TOP_THREAT_TYPES_SET = new Set(TOP_THREAT_TYPES);
@@ -78,7 +104,7 @@ function extractFeatures(result, meta) {
   const distinctTypes = new Set(threats.map(t => t.type));
   features.distinct_threat_types = distinctTypes.size;
 
-  // --- Per-type counts (top 31 types) ---
+  // --- Per-type counts (top 47 types) ---
   const typeCounts = Object.create(null);
   for (const t of threats) {
     typeCounts[t.type] = (typeCounts[t.type] || 0) + 1;