diff --git a/nanobot/channels/websocket.py b/nanobot/channels/websocket.py
index 1622a37d..47c79e87 100644
--- a/nanobot/channels/websocket.py
+++ b/nanobot/channels/websocket.py
@@ -745,6 +745,9 @@ class WebSocketChannel(BaseChannel):
             content_type=mime,
             extra_headers=[
                 ("Cache-Control", "private, max-age=31536000, immutable"),
+                # Paired with the MIME whitelist above: prevents browsers from
+                # MIME-sniffing an octet-stream fallback into executable HTML.
+                ("X-Content-Type-Options", "nosniff"),
             ],
         )
 
@@ -942,6 +945,8 @@ class WebSocketChannel(BaseChannel):
         Returns ``(paths, None)`` on success or ``([], reason)`` on the first
         failure — the caller is expected to surface ``reason`` to the client
         and skip publishing so no half-formed message ever reaches the agent.
+        On failure, any images already written to disk earlier in the same
+        call are unlinked so partial ingress doesn't leak orphan files.
         ``reason`` is a short, stable token suitable for UI localization.
 
         Shape: ``list[{"data_url": str, "name"?: str | None}]``.
@@ -950,28 +955,39 @@ class WebSocketChannel(BaseChannel):
             return [], "too_many_images"
         media_dir = get_media_dir("websocket")
         paths: list[str] = []
+
+        def _abort(reason: str) -> tuple[list[str], str]:
+            for p in paths:
+                try:
+                    Path(p).unlink(missing_ok=True)
+                except OSError as exc:
+                    logger.warning(
+                        "websocket: failed to unlink partial media {}: {}", p, exc
+                    )
+            return [], reason
+
         for item in media:
             if not isinstance(item, dict):
-                return [], "malformed"
+                return _abort("malformed")
             data_url = item.get("data_url")
             if not isinstance(data_url, str) or not data_url:
-                return [], "malformed"
+                return _abort("malformed")
             mime = _extract_data_url_mime(data_url)
             if mime is None:
-                return [], "decode"
+                return _abort("decode")
             if mime not in _IMAGE_MIME_ALLOWED:
-                return [], "mime"
+                return _abort("mime")
             try:
                 saved = save_base64_data_url(
                     data_url, media_dir, max_bytes=_MAX_IMAGE_BYTES,
                 )
             except FileSizeExceeded:
-                return [], "size"
+                return _abort("size")
             except Exception as exc:
                 logger.warning("websocket: media decode failed: {}", exc)
-                return [], "decode"
+                return _abort("decode")
             if saved is None:
-                return [], "decode"
+                return _abort("decode")
             paths.append(saved)
         return paths, None
 
diff --git a/tests/channels/test_websocket_envelope_media.py b/tests/channels/test_websocket_envelope_media.py
index 4fe5f642..97540804 100644
--- a/tests/channels/test_websocket_envelope_media.py
+++ b/tests/channels/test_websocket_envelope_media.py
@@ -354,7 +354,11 @@ async def test_message_rejected_when_media_field_is_not_list() -> None:
 
 @pytest.mark.asyncio
 async def test_failed_media_does_not_partially_persist(tmp_path) -> None:
-    """If the second image is invalid, the first must not be forwarded."""
+    """If the second image is invalid, the first must not be forwarded.
+
+    Also: images already written in this call are cleaned up on failure, so
+    a mixed-valid/invalid batch never leaves orphan files in the media dir.
+    """
     channel = _make_channel()
     mock_conn = AsyncMock()
     envelope = {
@@ -373,11 +377,11 @@ async def test_failed_media_does_not_partially_persist(tmp_path) -> None:
         await channel._dispatch_envelope(mock_conn, "client-1", envelope)
 
     channel._handle_message.assert_not_awaited()
-    # The first image was saved to disk (we don't roll it back — the caller
-    # is expected to not reference it) but the agent never sees the paths.
-    # That's the important invariant: no partial publish.
     err = json.loads(mock_conn.send.call_args[0][0])
     assert err["reason"] == "mime"
+    # Partial-batch failures must not leak files to disk.
+    leftover = [p for p in tmp_path.iterdir() if p.is_file()]
+    assert leftover == [], f"orphan media after rejected batch: {leftover}"
 
 
 @pytest.mark.asyncio
diff --git a/tests/channels/test_websocket_media_route.py b/tests/channels/test_websocket_media_route.py
index 14577355..0e08dc14 100644
--- a/tests/channels/test_websocket_media_route.py
+++ b/tests/channels/test_websocket_media_route.py
@@ -155,6 +155,8 @@ async def test_media_route_serves_signed_file(
     assert resp.headers["content-type"].startswith("image/png")
     # Immutable cache header lets the browser skip round-trips on replay.
     assert "immutable" in resp.headers.get("cache-control", "")
+    # nosniff keeps the browser from second-guessing our Content-Type.
+    assert resp.headers.get("x-content-type-options") == "nosniff"
 
 
 @pytest.mark.asyncio
@@ -281,6 +283,9 @@ async def test_media_route_degrades_non_image_to_octet_stream(
             await server_task
     assert resp.status_code == 200
     assert resp.headers["content-type"].startswith("application/octet-stream")
+    # nosniff is the actual defence when we downgrade to octet-stream:
+    # without it the browser might still sniff the bytes as HTML.
+    assert resp.headers.get("x-content-type-options") == "nosniff"
 
 
 # ---------------------------------------------------------------------------