guard uninitialized count from descr sscanf in npy parser

test/correctness/image_io.cpp

@@ -294,6 +294,38 @@ void test_read_big_endian_row_channel_offset() {
     }
 }
 
+// A .npy whose 'descr' is missing its closing quote used to walk the header
+// parser's cursor off the end of the buffer (sscanf matched %c%c%d and returned
+// 3, but the %n was never reached, leaving the byte count uninitialized). The
+// loader should just reject the file.
+void test_malformed_npy_header() {
+    std::ostringstream o;
+    o << Internal::get_test_tmp_dir() << "malformed_descr.npy";
+    std::string filename = o.str();
+
+    // v1 header: magic, version 1.0, 2-byte header length, then the dict.
+    // (6 + 2 + 2 + header_len) must be a multiple of 64.
+    std::string dict = "{'descr': '<f8";  // no closing quote after the type
+    const size_t header_len = 54;
+    dict.resize(header_len, ' ');
+
+    std::ofstream fs(filename.c_str(), std::ofstream::binary);
+    const char magic[8] = {'\x93', 'N', 'U', 'M', 'P', 'Y', '\x01', '\x00'};
+    fs.write(magic, 8);
+    const char len_le[2] = {(char)(header_len & 0xff), (char)((header_len >> 8) & 0xff)};
+    fs.write(len_le, 2);
+    fs.write(dict.data(), dict.size());
+    const std::vector<char> payload(64, 0);
+    fs.write(payload.data(), payload.size());
+    fs.close();
+
+    Buffer<> im;
+    if (Tools::load<Buffer<>>(filename, &im)) {
+        std::cout << "Malformed .npy header was accepted by the loader\n";
+        exit(1);
+    }
+}
+
 #ifndef HALIDE_NO_PNG
 void test_png_unsupported_bit_depth() {
     // A 1-bit grayscale PNG is a valid file, but load_png only supports 8- and
@@ -352,6 +384,7 @@ int main(int argc, char **argv) {
     do_test<double>();
     test_mat_header();
     test_read_big_endian_row_channel_offset();
+    test_malformed_npy_header();
 #ifndef HALIDE_NO_PNG
     test_png_unsupported_bit_depth();
 #endif

tools/halide_image_io.h

@@ -1236,8 +1236,12 @@ struct NpyHeader {
         }
         while (true) {
             char endian;
-            int consumed;
-            if (std::sscanf(ptr, "'descr': '%c%c%d'%n", &endian, &type_code, &type_bytes, &consumed) == 3) {
+            int consumed = 0;
+            // %n is not counted in sscanf's return value and is only reached
+            // if the trailing quote matches, so a descr missing its closing
+            // quote returns 3 with consumed still 0. Requiring consumed > 0
+            // avoids advancing ptr by an indeterminate amount.
+            if (std::sscanf(ptr, "'descr': '%c%c%d'%n", &endian, &type_code, &type_bytes, &consumed) == 3 && consumed > 0) {
                 if (endian != '<' && endian != '|') {
                     return false;
                 }