@lateos/npm-scan 0.15.2 → 0.15.3

package/README.md

@@ -3,7 +3,7 @@
 [![npm version](https://img.shields.io/npm/v/@lateos/npm-scan?style=flat-square)](https://www.npmjs.com/package/@lateos/npm-scan)
 [![License](https://img.shields.io/badge/license-Apache%202.0%20%2B%20Commons%20Clause-blue?style=flat-square)](LICENSING.md)
 [![Node](https://img.shields.io/badge/node-%3E%3D18-brightgreen?style=flat-square)](package.json)
-[![Tests](https://img.shields.io/badge/tests-371%20passing-brightgreen?style=flat-square)](https://github.com/lateos-ai/npm-scan)
+[![Tests](https://img.shields.io/badge/tests-384%20passing-brightgreen?style=flat-square)](https://github.com/lateos-ai/npm-scan)
 [![Coverage](https://img.shields.io/badge/coverage-90%25-brightgreen?style=flat-square)](https://github.com/lateos-ai/npm-scan)
 [![Docker](https://img.shields.io/badge/docker-lateos%2Fnpm--scan-2496ED?style=flat-square&logo=docker)](https://hub.docker.com/r/lateos/npm-scan)
 [![Sigstore](https://img.shields.io/static/v1?label=Sigstore&message=Provenance&color=green&style=flat-square&logo=sigstore)](https://github.com/lateos-ai/npm-scan/actions/workflows/publish.yml)
@@ -24,6 +24,8 @@ The 2025–2026 wave of npm supply chain attacks proved that traditional tooling
 
 Attackers have moved past simple typosquatting. They now ship **obfuscated preinstall hooks**, **credential harvesters hidden behind environment detection**, **dormant backdoors with time-based activation**, and **worm-style transitive propagation** that spreads through peer dependencies.
 
+A growing attack vector is **HuggingFace org impersonation** — packages that masquerade as legitimate HF model repositories (e.g., `0penai/gpt2` instead of `openai/gpt2`) to trick users into downloading malicious model artifacts during CI/CD pipelines, often bundled with suspicious binaries (`.exe`, `.dll`) in model repos that deep-learned tools trust by default.
+
 The **Megalodon campaign** (2026) alone compromised 5,500+ repositories via fake GitHub PRs, malicious workflow injection, and cloud credential exfiltration — all coordinated through a single actor automating the entire kill chain. **@lateos/npm-scan** now detects artifacts of this campaign out of the box.
 
 **npm audit** checks known CVEs. **Snyk** scans for vulnerabilities. **Socket** looks at package behavior. None of them were designed for the generation of attacks that emerged in 2025 — attacks that look benign until they reach production.
@@ -45,6 +47,7 @@ The **Megalodon campaign** (2026) alone compromised 5,500+ repositories via fake
 | Sandbox evasion detection (ATK-010) | ❌ | ❌ | ❌ | ✅ |
 | Transitive worm propagation (ATK-011) | ❌ | ❌ | ❌ | ✅ |
 | Campaign detection (Megalodon CI/CD) | ❌ | ❌ | ❌ | ✅ |
+| HF model repo impersonation + README clone | ❌ | ❌ | ❌ | ✅ |
 | Attack taxonomy (ATK series) | ❌ | ❌ | ❌ | ✅ |
 | SBOM output (CycloneDX + SPDX) | ❌ | ✅ | ❌ | ✅ |
 | SARIF v2.1 (GitHub Code Scanning) | ❌ | ❌ | ❌ | ✅ |
@@ -74,6 +77,7 @@ The **Megalodon campaign** (2026) alone compromised 5,500+ repositories via fake
 | 🛡️ | **Zero telemetry** | No data leaves your machine. No cloud. No callbacks. |
 | 💾 | **Local scan history** | SQLite-backed persistence, zero external dependencies |
 | 🪝 | **Pre-commit hook** | Block threats before commit — one-liner install, scans `package-lock.json` changes |
+| 🤖 | **HF impersonation detection** | Detects typosquatted HuggingFace orgs (Jaro-Winkler), README clones (SimHash), artifact mismatches (`.exe` in model repos), and new-org amplifier — with lazy two-stage evaluation, zero network in Stage 1 |
 | 📎 | **Yarn + pnpm support** | `scan-lockfile` parses `yarn.lock` and `pnpm-lock.yaml` alongside `package-lock.json` |
 
 ---
@@ -283,9 +287,11 @@ npm-scan report --pdf             # all scans (premium)
 | **ATK-010** | Sandbox evasion / anti-analysis | Behavioral | 🟠 medium | SR-10.3 |
 | **ATK-011** | Transitive propagation (worm-style lateral spread) | Behavioral | 🔴 high | SR-11.4 |
 | **MEGALODON** | Megalodon CI/CD campaign — workflow C2 exfil, credential harvest, publish velocity spike, publisher drift | Static + Registry | ⚫ critical | SR-3.1, SR-7.5 |
+| **HF_IMPERSONATION** | HuggingFace org spoof detection — Jaro-Winkler similarity against 15 known-good orgs, SimHash README clone detection, artifact mismatch (`.exe`/`.dll` in model repos), postinstall escalation, new-org amplifier | Static + Network (Stage 2) | 🔴 high / ⚫ critical | SR-2.1 |
 
 > **How evasive attacks are caught:** ATK-009 detects packages that check `process.env.CI`, probe hostnames, or use time-based activation. ATK-010 flags `debugger` statements, `os.hostname()` probes, and env fingerprinting. ATK-011 traces peer dependency graphs to detect worm-like propagation patterns.  
 > **MEGALODON** campaign detection analyzes bundled `.github/workflows/` files for C2 co-occurrence and base64 decode chains, scans tarball files for credential + outbound network patterns, detects version publish velocity spikes via npm registry metadata, and identifies publisher account drift — all without any network calls beyond the initial package fetch.  
+> **HF_IMPERSONATION** detection uses a lazy two-stage evaluation: Stage 1 scans `package.json` scripts and JS/TS sources for HuggingFace references (URLs, `from_pretrained()`, `hub.download()`) and runs Jaro-Winkler similarity against 15 known-good HF orgs — zero network. If spoofs are found, Stage 2 fetches the HF model API, computes SimHash of both READMEs for clone detection, validates artifact type consistency (e.g., `transformers` library with `.exe` files is flagged as critical), applies a new-org amplifier (<30 days), and escalates when the reference appears in a lifecycle script.  
 > See [`docs/attack-taxonomy.md`](docs/attack-taxonomy.md) for full evasion surface documentation and PoC examples.
 
 ---
@@ -632,7 +638,7 @@ See the [Docker quick-start section](#-run-lateosnpm-scan-anywhere-with-docker--
 
 ### Free tier (shipped)
 
-- All 11 ATK detectors + **MEGALODON** CI/CD campaign detection (D1–D6)
+- All 11 ATK detectors + **MEGALODON** CI/CD campaign detection (D1–D6) + **HF_IMPERSONATION** detector
 - SBOM output (CycloneDX + SPDX)
 - HTML, text, and compliance reports (NIST + EU CRA)
 - Policy-as-code engine (YAML)
@@ -701,6 +707,7 @@ node --test test/detectors-corpus.test.js
 - `test/report-snapshots.test.js` — HTML/text/CRA/PDF format assertions
 - `test/report.test.js` — SARIF, CSV, STIG, risk score format tests
 - `test/lockfile.test.js` — npm/yarn/pnpm parser, auto-detect, ATK-007/011 lockfile tests
+- `test/hf-impersonation.test.js` — 13 HF impersonation detection tests (no-ref, exact match, spoof, README clone, artifact mismatch, postinstall escalation, new-org tag)
 - `test/cli.test.js` — commander integration tests (help, version, scan, report, error handling)
 - `test/cli-lockfile.test.js` — scan-lockfile CLI options, yarn/pnpm/monorepo/watch tests
 

package/backend/detectors/hf-impersonation/index.js

@@ -0,0 +1,396 @@
+import { KNOWN_HF_ORGS } from './known-orgs.js';
+import { jaroWinkler } from './jaro-winkler.js';
+import { simhash, similarity as simhashSimilarity } from './simhash.js';
+
+const HF_URL_PATTERN = /(?:huggingface\.co|hf\.co)\/([^\/\s"'>]+)\/([^\/\s"'>]+)/g;
+const FROM_PRETRAINED_PATTERN = /from_pretrained\(\s*["']([^"']+\/[^"']+)["']/g;
+const HUB_DOWNLOAD_SINGLE = /hub\.download\(\s*["']([^"']+\/[^"']+)["']/g;
+const HUB_DOWNLOAD_DOUBLE = /hub\.download\(\s*["']([^"']+)["']\s*,\s*["']([^"']+)["']/g;
+
+const LIFECYCLE_SCRIPTS = new Set(['postinstall', 'prepare', 'install']);
+const API_BASE = 'https://huggingface.co';
+
+const SEVERITY_SCORE = { none: 0, low: 1, medium: 2, high: 3, critical: 4 };
+const SEVERITY_LABELS = ['none', 'low', 'medium', 'high', 'critical'];
+
+const HF_ARTIFACT_LIBS = new Set(['transformers', 'diffusers', 'sentence-transformers', 'gguf', 'safetensors']);
+const SUSPICIOUS_EXTENSIONS = /\.(exe|msi|bat|ps1|dll)$/i;
+
+const _cache = new Map();
+const CACHE_TTL = 3600 * 1000;
+let _lastFetchTime = 0;
+
+function severityIndex(sev) {
+  return SEVERITY_SCORE[sev] || 0;
+}
+
+function maxSeverity(a, b) {
+  return severityIndex(a) >= severityIndex(b) ? a : b;
+}
+
+function sleep(ms) {
+  return new Promise(r => setTimeout(r, ms));
+}
+
+async function fetchWithCache(url) {
+  const cached = _cache.get(url);
+  if (cached && Date.now() - cached.fetchedAt < CACHE_TTL) {
+    return cached.data;
+  }
+  const now = Date.now();
+  const elapsed = now - _lastFetchTime;
+  if (elapsed < 100) {
+    await sleep(100 - elapsed);
+  }
+  _lastFetchTime = Date.now();
+  let res;
+  try {
+    res = await fetch(url);
+    if (res.status === 429) {
+      const retryAfter = parseInt(res.headers.get('Retry-After') || '5', 10);
+      await sleep(retryAfter * 1000);
+      res = await fetch(url);
+    }
+    if (!res.ok) {
+      console.debug(`HF API warning: ${url} returned ${res.status}`);
+      return null;
+    }
+    const data = await res.json();
+    _cache.set(url, { data, fetchedAt: Date.now() });
+    return data;
+  } catch (err) {
+    console.debug(`HF API warning: ${err.message}`);
+    return null;
+  }
+}
+
+async function fetchReadme(url) {
+  const cached = _cache.get(url);
+  if (cached && Date.now() - cached.fetchedAt < CACHE_TTL) {
+    return cached.data;
+  }
+  const now = Date.now();
+  const elapsed = now - _lastFetchTime;
+  if (elapsed < 100) {
+    await sleep(100 - elapsed);
+  }
+  _lastFetchTime = Date.now();
+  try {
+    const res = await fetch(url);
+    if (res.status === 429) {
+      const retryAfter = parseInt(res.headers.get('Retry-After') || '5', 10);
+      await sleep(retryAfter * 1000);
+      const retryRes = await fetch(url);
+      if (!retryRes.ok) return null;
+      const text = await retryRes.text();
+      _cache.set(url, { data: text, fetchedAt: Date.now() });
+      return text;
+    }
+    if (!res.ok) return null;
+    const text = await res.text();
+    _cache.set(url, { data: text, fetchedAt: Date.now() });
+    return text;
+  } catch (err) {
+    console.debug(`HF README warning: ${err.message}`);
+    return null;
+  }
+}
+
+function findClosestOrg(spoofedOrg) {
+  const lowerOrg = String(spoofedOrg).toLowerCase();
+  let best = { org: null, score: 0 };
+  for (const known of KNOWN_HF_ORGS) {
+    const score = jaroWinkler(lowerOrg, known.toLowerCase());
+    if (score >= 0.82 && score > best.score) {
+      best = { org: known, score };
+    }
+  }
+  return best;
+}
+
+function extractHFTuples(pkgJson, allFiles) {
+  const tuples = new Set();
+  let postinstallFetchFlag = false;
+
+  const scripts = pkgJson?.scripts || {};
+  let m;
+  for (const [hook, script] of Object.entries(scripts)) {
+    if (typeof script !== 'string') continue;
+
+    HF_URL_PATTERN.lastIndex = 0;
+    while ((m = HF_URL_PATTERN.exec(script)) !== null) {
+      tuples.add(`${m[1]}/${m[2]}`);
+      if (LIFECYCLE_SCRIPTS.has(hook)) {
+        postinstallFetchFlag = true;
+      }
+    }
+
+    FROM_PRETRAINED_PATTERN.lastIndex = 0;
+    while ((m = FROM_PRETRAINED_PATTERN.exec(script)) !== null) {
+      tuples.add(m[1]);
+      if (LIFECYCLE_SCRIPTS.has(hook)) {
+        postinstallFetchFlag = true;
+      }
+    }
+
+    HUB_DOWNLOAD_SINGLE.lastIndex = 0;
+    while ((m = HUB_DOWNLOAD_SINGLE.exec(script)) !== null) {
+      tuples.add(m[1]);
+      if (LIFECYCLE_SCRIPTS.has(hook)) {
+        postinstallFetchFlag = true;
+      }
+    }
+
+    HUB_DOWNLOAD_DOUBLE.lastIndex = 0;
+    while ((m = HUB_DOWNLOAD_DOUBLE.exec(script)) !== null) {
+      tuples.add(`${m[1]}/${m[2]}`);
+      if (LIFECYCLE_SCRIPTS.has(hook)) {
+        postinstallFetchFlag = true;
+      }
+    }
+  }
+
+  if (allFiles) {
+    for (const file of allFiles) {
+      if (!file.path?.match(/\.(js|ts|jsx|tsx|mjs|cjs)$/i)) continue;
+      const content = typeof file.content === 'string' ? file.content : '';
+
+      HF_URL_PATTERN.lastIndex = 0;
+      while ((m = HF_URL_PATTERN.exec(content)) !== null) {
+        tuples.add(`${m[1]}/${m[2]}`);
+      }
+
+      FROM_PRETRAINED_PATTERN.lastIndex = 0;
+      while ((m = FROM_PRETRAINED_PATTERN.exec(content)) !== null) {
+        tuples.add(m[1]);
+      }
+
+      HUB_DOWNLOAD_SINGLE.lastIndex = 0;
+      while ((m = HUB_DOWNLOAD_SINGLE.exec(content)) !== null) {
+        tuples.add(m[1]);
+      }
+
+      HUB_DOWNLOAD_DOUBLE.lastIndex = 0;
+      while ((m = HUB_DOWNLOAD_DOUBLE.exec(content)) !== null) {
+        tuples.add(`${m[1]}/${m[2]}`);
+      }
+    }
+  }
+
+  return { tuples, postinstallFetchFlag };
+}
+
+function buildHFOrgSpoofFinding(referencedRepo, org, canonicalOrg, similarityScore, postinstallFetchFlag, tags, hfMeta) {
+  const finding = {
+    id: 'HF_ORG_SPOOF',
+    severity: 'high',
+    title: 'HuggingFace org impersonation',
+    description: `Repository "${referencedRepo}" references org "${org}" which is similar to known HF org "${canonicalOrg.org}" (similarity: ${similarityScore.toFixed(3)})`,
+    evidence: JSON.stringify({
+      referencedRepo,
+      canonicalOrg: canonicalOrg.org,
+      similarityScore,
+      tags: tags || [],
+    }),
+    referencedRepo,
+    canonicalOrg: canonicalOrg.org,
+    similarityScore,
+    tags: tags || [],
+    ipiClass: 'SUPPLY_CHAIN',
+  };
+  if (hfMeta) {
+    finding.hfMeta = hfMeta;
+  }
+  return finding;
+}
+
+async function runStage2(spoofFindings, orgsToCheck, postinstallFetchFlag) {
+  const newFindings = [];
+
+  for (const [referencedRepo, { org, canonicalOrg, similarityScore, finding }] of orgsToCheck) {
+    const tags = [];
+    let hfMeta = null;
+
+    const modelUrl = `${API_BASE}/api/models/${referencedRepo}`;
+    const canonicalUrl = canonicalOrg.org !== org ? `${API_BASE}/api/models/${canonicalOrg.org}/${referencedRepo.split('/')[1]}` : null;
+    const userUrl = `${API_BASE}/api/users/${org}`;
+
+    const spoofedModel = await fetchWithCache(modelUrl);
+    const canonicalModel = canonicalUrl ? await fetchWithCache(canonicalUrl) : null;
+    const userData = await fetchWithCache(userUrl);
+
+    // Org age check for NEW_ORG tag
+    if (userData?.dateCreated) {
+      const created = new Date(userData.dateCreated);
+      const ageDays = (Date.now() - created.getTime()) / (1000 * 60 * 60 * 24);
+      hfMeta = {
+        orgAgeDays: Math.round(ageDays),
+        repoDownloads: spoofedModel?.downloads ?? 0,
+      };
+      if (ageDays < 30) {
+        tags.push('NEW_ORG');
+      }
+    }
+
+    // README clone check
+    if (canonicalOrg.org !== org) {
+      const readmeSpoof = await fetchReadme(`${API_BASE}/${referencedRepo}/resolve/main/README.md`);
+      const readmeCanonical = await fetchReadme(`${API_BASE}/${canonicalOrg.org}/${referencedRepo.split('/')[1]}/resolve/main/README.md`);
+
+      if (readmeSpoof && readmeCanonical) {
+        const fp1 = simhash(readmeSpoof);
+        const fp2 = simhash(readmeCanonical);
+        const simScore = simhashSimilarity(fp1, fp2);
+
+        if (simScore >= 0.9) {
+          const readmeFinding = {
+            id: 'HF_README_CLONE',
+            severity: 'high',
+            title: 'HuggingFace README clone',
+            description: `README of "${referencedRepo}" is highly similar (${(simScore * 100).toFixed(1)}%) to canonical org "${canonicalOrg.org}/${referencedRepo.split('/')[1]}"`,
+            evidence: JSON.stringify({
+              referencedRepo,
+              canonicalOrg: canonicalOrg.org,
+              similarityScore: simScore,
+              tags: [],
+            }),
+            referencedRepo,
+            canonicalOrg: canonicalOrg.org,
+            similarityScore: simScore,
+            tags: [],
+            ipiClass: 'SUPPLY_CHAIN',
+          };
+          if (hfMeta) readmeFinding.hfMeta = hfMeta;
+          newFindings.push(readmeFinding);
+        }
+      }
+    }
+
+    // Artifact mismatch check
+    if (spoofedModel?.cardData?.library_name && spoofedModel?.siblings) {
+      const libName = spoofedModel.cardData.library_name;
+      if (HF_ARTIFACT_LIBS.has(libName)) {
+        for (const sibling of spoofedModel.siblings) {
+          const fn = sibling.rfilename || '';
+          if (SUSPICIOUS_EXTENSIONS.test(fn)) {
+            const artifactFinding = {
+              id: 'HF_ARTIFACT_MISMATCH',
+              severity: 'critical',
+              title: 'HF artifact mismatch — suspicious binary in model repo',
+              description: `Model "${referencedRepo}" declares library "${libName}" but contains suspicious file "${fn}"`,
+              evidence: JSON.stringify({
+                referencedRepo,
+                artifactConflict: { declaredType: libName, suspiciousFilename: fn },
+                tags: [],
+              }),
+              referencedRepo,
+              artifactConflict: { declaredType: libName, suspiciousFilename: fn },
+              tags: [],
+              ipiClass: 'SUPPLY_CHAIN',
+            };
+            if (hfMeta) artifactFinding.hfMeta = hfMeta;
+            newFindings.push(artifactFinding);
+            break;
+          }
+        }
+      }
+    }
+
+    // Apply NEW_ORG and POSTINSTALL_FETCH tags to all findings for this repo
+    const repoSpoofFindings = spoofFindings.filter(f => f.referencedRepo === referencedRepo);
+    for (const sf of repoSpoofFindings) {
+      if (tags.length > 0) {
+        if (!sf.tags) sf.tags = [];
+        for (const t of tags) {
+          if (!sf.tags.includes(t)) sf.tags.push(t);
+        }
+      }
+      if (hfMeta) {
+        sf.hfMeta = hfMeta;
+      }
+    }
+    for (const nf of newFindings) {
+      if (nf.referencedRepo === referencedRepo) {
+        if (tags.length > 0) {
+          if (!nf.tags) nf.tags = [];
+          for (const t of tags) {
+            if (!nf.tags.includes(t)) nf.tags.push(t);
+          }
+        }
+      }
+    }
+  }
+
+  // POSTINSTALL_FETCH escalation
+  if (postinstallFetchFlag) {
+    const allStage2Findings = [...spoofFindings, ...newFindings];
+    const escalatedRepos = new Set();
+    for (const f of allStage2Findings) {
+      if (f.referencedRepo) escalatedRepos.add(f.referencedRepo);
+    }
+    for (const f of allStage2Findings) {
+      if (escalatedRepos.has(f.referencedRepo)) {
+        if (severityIndex(f.severity) < severityIndex('critical')) {
+          f.severity = 'critical';
+        }
+        if (!f.tags) f.tags = [];
+        if (!f.tags.includes('POSTINSTALL_ESCALATED')) {
+          f.tags.push('POSTINSTALL_ESCALATED');
+        }
+      }
+    }
+  }
+
+  return newFindings;
+}
+
+export async function scan(pkgJson, files = [], registryMeta = null, allFiles = null) {
+  const { tuples, postinstallFetchFlag } = extractHFTuples(pkgJson, allFiles || files);
+
+  if (tuples.size === 0) return [];
+
+  // Stage 1: org spoof detection (local only)
+  const spoofFindings = [];
+  const orgsToCheck = []; // [referencedRepo, { org, canonicalOrg, similarityScore, finding }]
+
+  for (const tuple of tuples) {
+    const parts = tuple.split('/');
+    if (parts.length < 2) continue;
+    const org = parts[0];
+
+    const canonicalOrg = findClosestOrg(org);
+    if (!canonicalOrg.org) continue;
+    if (org.toLowerCase() === canonicalOrg.org.toLowerCase()) continue;
+
+    const finding = buildHFOrgSpoofFinding(tuple, org, canonicalOrg, canonicalOrg.score, postinstallFetchFlag, []);
+    spoofFindings.push(finding);
+    orgsToCheck.push([tuple, { org, canonicalOrg, similarityScore: canonicalOrg.score, finding }]);
+  }
+
+  if (spoofFindings.length === 0) return [];
+
+  // Stage 2: network checks
+  const stage2Findings = await runStage2(spoofFindings, orgsToCheck, postinstallFetchFlag);
+
+  // Deduplicate POSTINSTALL_ESCALATED tag in evidence
+  for (const f of [...spoofFindings, ...stage2Findings]) {
+    if (f.tags && f.tags.length > 0) {
+      try {
+        const ev = JSON.parse(f.evidence);
+        ev.tags = [...f.tags];
+        f.evidence = JSON.stringify(ev);
+      } catch {
+        // evidence wasn't JSON, leave as-is
+      }
+    }
+  }
+
+  return [...spoofFindings, ...stage2Findings];
+}
+
+export function clearCache() {
+  _cache.clear();
+  _lastFetchTime = 0;
+}
+
+export { KNOWN_HF_ORGS, jaroWinkler, simhash, simhashSimilarity };

package/backend/detectors/hf-impersonation/jaro-winkler.js

@@ -0,0 +1,44 @@
+export function jaroWinkler(s1, s2) {
+  if (s1 === s2) return 1;
+  const len1 = s1.length, len2 = s2.length;
+  if (len1 === 0 || len2 === 0) return 0;
+
+  const matchDist = Math.floor(Math.max(len1, len2) / 2) - 1;
+  const matches1 = new Array(len1).fill(false);
+  const matches2 = new Array(len2).fill(false);
+  let matches = 0;
+
+  for (let i = 0; i < len1; i++) {
+    const start = Math.max(0, i - matchDist);
+    const end = Math.min(len2, i + matchDist + 1);
+    for (let j = start; j < end; j++) {
+      if (matches2[j]) continue;
+      if (s1[i] !== s2[j]) continue;
+      matches1[i] = true;
+      matches2[j] = true;
+      matches++;
+      break;
+    }
+  }
+
+  if (matches === 0) return 0;
+
+  let transpositions = 0, k = 0;
+  for (let i = 0; i < len1; i++) {
+    if (!matches1[i]) continue;
+    while (!matches2[k]) k++;
+    if (s1[i] !== s2[k]) transpositions++;
+    k++;
+  }
+
+  const jaro = (matches / len1 + matches / len2 + (matches - transpositions / 2) / matches) / 3;
+
+  let prefix = 0;
+  const maxPrefix = Math.min(4, len1, len2);
+  for (let i = 0; i < maxPrefix; i++) {
+    if (s1[i] === s2[i]) prefix++;
+    else break;
+  }
+
+  return jaro + prefix * 0.1 * (1 - jaro);
+}

package/backend/detectors/hf-impersonation/known-orgs.js

@@ -0,0 +1,5 @@
+export const KNOWN_HF_ORGS = [
+  'openai', 'meta-llama', 'mistralai', 'google', 'microsoft',
+  'stabilityai', 'EleutherAI', 'huggingface', 'tiiuae', 'cohere',
+  'anthropic', 'deepseek-ai', 'Qwen', 'NousResearch', 'teknium',
+];

package/backend/detectors/hf-impersonation/simhash.js

@@ -0,0 +1,46 @@
+function hashToken(str) {
+  let hash = 5381;
+  for (let i = 0; i < str.length; i++) {
+    hash = ((hash << 5) + hash) + str.charCodeAt(i);
+    hash = hash & hash;
+  }
+  return hash >>> 0;
+}
+
+export function simhash(text) {
+  const v = new Array(64).fill(0);
+  const tokens = text.toLowerCase().split(/\s+/).filter(Boolean);
+
+  for (const token of tokens) {
+    const h = hashToken(token);
+    for (let i = 0; i < 64; i++) {
+      if ((h >> i) & 1) {
+        v[i] += 1;
+      } else {
+        v[i] -= 1;
+      }
+    }
+  }
+
+  let fingerprint = 0n;
+  for (let i = 0; i < 64; i++) {
+    if (v[i] > 0) {
+      fingerprint |= (1n << BigInt(i));
+    }
+  }
+  return fingerprint;
+}
+
+export function hammingDistance(a, b) {
+  let xor = a ^ b;
+  let count = 0;
+  while (xor > 0n) {
+    count += Number(xor & 1n);
+    xor >>= 1n;
+  }
+  return count;
+}
+
+export function similarity(a, b) {
+  return 1 - hammingDistance(a, b) / 64;
+}

package/backend/detectors/index.js

@@ -10,6 +10,7 @@ import * as atk009 from './atk-009-dormant-trigger.js';
 import * as atk010 from './atk-010-sandbox-evasion.js';
 import * as atk011 from './atk-011-transitive-prop.js';
 import { scanAll as megalodonScan } from './megalodon/index.js';
+import { scan as hfScan } from './hf-impersonation/index.js';
 
 export async function runAll(pkgJson, files = [], registryMeta = null, allFiles = null) {
   const findings = [];
@@ -25,5 +26,6 @@ export async function runAll(pkgJson, files = [], registryMeta = null, allFiles
   findings.push(...await atk010.scan(pkgJson, files));
   findings.push(...await atk011.scan(pkgJson, files));
   findings.push(...await megalodonScan(pkgJson, allFiles || files, registryMeta));
+  findings.push(...await hfScan(pkgJson, files, registryMeta, allFiles || files));
   return findings.sort((a, b) => b.severity.localeCompare(a.severity));
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lateos/npm-scan",
-  "version": "0.15.2",
+  "version": "0.15.3",
   "description": "Modern npm supply chain security scanner — detects obfuscated payloads, credential stealers, conditional triggers, sandbox evasion, and worm-like propagation. 11 attack types, SBOM, NIST/EU CRA compliance reporting.",
   "main": "backend/index.js",
   "bin": {