fix(upload): validate audio magic bytes in addition to MIME type

client/src/hooks/useSpeechHistory.js

@@ -136,12 +136,6 @@ const addMessage = useCallback((text) => {
     // Check existing message
     const existing = prev.find((m) => m.text === trimmed);
 
-    // Preserve existing ID if duplicate found
-    const entry = existing || {
-      id: crypto.randomUUID(),
-      text: trimmed,
-      timestamp,
-    };
     // Preserve existing ID if duplicate found, but update timestamp
     // so re-spoken messages sort correctly after a page reload.
     const entry = existing

server/controllers/voiceController.js

@@ -1,6 +1,7 @@
 // Implements ElevenLabs voice cloning and text-to-speech proxy handlers.
 import crypto from "crypto";
-import { getIsMock } from "../utils/mock.js"; // adjust path to actual location
+import { isValidAudioBuffer } from "../middleware/upload.js";
+import { getIsMock } from "../utils/mock.js";
 import { isValidLanguageCode } from "../utils/languages.js";
 
 const ELEVENLABS_BASE_URL = "https://api.elevenlabs.io/v1";
@@ -130,6 +131,14 @@ export async function cloneVoice(request, response, next) {
       return;
     }
 
+    // The MIME type checked in upload.js comes from the client and can be
+    // spoofed. Verify the buffer begins with a known audio magic-byte
+    // signature so arbitrary binary data cannot be forwarded to ElevenLabs.
+    if (!isValidAudioBuffer(audioFile.buffer)) {
+      response.status(400).json({ error: "Uploaded file does not appear to be valid audio." });
+      return;
+    }
+
     // --- mock mode: return a deterministic fixture voice_id ---
     if (getIsMock()) {
       console.warn("[VoiceForge] MOCK_ELEVENLABS: skipping real voice clone, returning fixture.");

server/middleware/upload.js

@@ -10,6 +10,39 @@ const ALLOWED_MIME_TYPES = [
   "audio/flac"
 ];
 
+// Known magic-byte signatures for audio formats accepted by ElevenLabs.
+// Each entry is { offset, bytes } where bytes is a Buffer to match at that
+// position in the uploaded file.
+const AUDIO_SIGNATURES = [
+  // WebM / Matroska  (EBML header: 0x1A 0x45 0xDF 0xA3)
+  { offset: 0, bytes: Buffer.from([0x1a, 0x45, 0xdf, 0xa3]) },
+  // WAV              (RIFF....WAVE)
+  { offset: 0, bytes: Buffer.from("RIFF", "ascii") },
+  // MP3 with ID3 tag
+  { offset: 0, bytes: Buffer.from("ID3", "ascii") },
+  // MP3 sync word    (0xFF 0xFB / 0xFF 0xFA / 0xFF 0xF3 and similar)
+  { offset: 0, bytes: Buffer.from([0xff, 0xfb]) },
+  { offset: 0, bytes: Buffer.from([0xff, 0xfa]) },
+  { offset: 0, bytes: Buffer.from([0xff, 0xf3]) },
+  { offset: 0, bytes: Buffer.from([0xff, 0xe3]) },
+  // OGG              (OggS)
+  { offset: 0, bytes: Buffer.from("OggS", "ascii") },
+  // FLAC
+  { offset: 0, bytes: Buffer.from("fLaC", "ascii") },
+  // AIFF
+  { offset: 0, bytes: Buffer.from("FORM", "ascii") },
+  // MP4 / M4A ftyp box (bytes 4-7 are "ftyp")
+  { offset: 4, bytes: Buffer.from("ftyp", "ascii") },
+];
+
+export function isValidAudioBuffer(buf) {
+  if (!buf || buf.length < 12) return false;
+  return AUDIO_SIGNATURES.some(({ offset, bytes }) => {
+    if (buf.length < offset + bytes.length) return false;
+    return buf.slice(offset, offset + bytes.length).equals(bytes);
+  });
+}
+
 const upload = multer({
   storage: multer.memoryStorage(),
   limits: {

server/test/voiceController.mock-mode.test.js

@@ -47,8 +47,12 @@ test("MOCK_ELEVENLABS: cloneVoice returns fixture voice_id without an API key",
   t.after(restore);
 
   const request = createRequest({ body: { name: "Contributor test voice" } });
+  // Buffer starting with WebM magic bytes (0x1A 0x45 0xDF 0xA3) followed by padding
   request.file = {
-    buffer: Buffer.from("fake-audio"),
+    buffer: Buffer.concat([
+      Buffer.from([0x1a, 0x45, 0xdf, 0xa3]),
+      Buffer.alloc(12)
+    ]),
     mimetype: "audio/webm",
     originalname: "test.webm"
   };
@@ -162,8 +166,12 @@ test("MOCK_ELEVENLABS is ignored in production: cloneVoice requires a real API k
   t.after(restore);
 
   const request = createRequest({ body: { name: "prod test" } });
+  // Buffer starting with WebM magic bytes (0x1A 0x45 0xDF 0xA3) followed by padding
   request.file = {
-    buffer: Buffer.from("fake-audio"),
+    buffer: Buffer.concat([
+      Buffer.from([0x1a, 0x45, 0xdf, 0xa3]),
+      Buffer.alloc(12)
+    ]),
     mimetype: "audio/webm",
     originalname: "test.webm"
   };