diff --git a/cpp/HybridTfliteModel.cpp b/cpp/HybridTfliteModel.cpp
index 1a7a3c41..78f0d42c 100644
--- a/cpp/HybridTfliteModel.cpp
+++ b/cpp/HybridTfliteModel.cpp
@@ -95,7 +95,17 @@ void HybridTfliteModel::copyInputBuffers(const std::vector<std::shared_ptr<Array
   for (int32_t i = 0; i < inputCount; i++) {
     TfLiteTensor* tensor = TfLiteInterpreterGetInputTensor(_interpreter, i);
     const std::shared_ptr<ArrayBuffer>& buffer = input[i];
-    TfLiteTensorCopyFromBuffer(tensor, buffer->data(), buffer->size());
+    TfLiteStatus status = TfLiteTensorCopyFromBuffer(tensor, buffer->data(), buffer->size());
+    if (status != kTfLiteOk) [[unlikely]] {
+      // TfLiteTensorCopyFromBuffer requires input_data_size == TfLiteTensorByteSize(tensor).
+      // On mismatch it leaves the tensor untouched, so without this check inference would
+      // silently run on the previous (or zero-initialized) contents of the tensor.
+      throw std::runtime_error("TFLite: Input buffer " + std::to_string(i) + " size (" +
+                               std::to_string(buffer->size()) + ") does not match input tensor \"" +
+                               std::string(TfLiteTensorName(tensor)) + "\" expected size (" +
+                               std::to_string(TfLiteTensorByteSize(tensor)) +
+                               ")! Status: " + tfLiteStatusToString(status));
+    }
   }
 }
 
diff --git a/example/__tests__/tflite.harness.ts b/example/__tests__/tflite.harness.ts
index d0c6854e..6ecf8ec7 100644
--- a/example/__tests__/tflite.harness.ts
+++ b/example/__tests__/tflite.harness.ts
@@ -69,6 +69,11 @@ function filledInputBuffer(input: Tensor, byte: number): ArrayBuffer {
   return buf;
 }
 
+/** Output buffers are pre-allocated per tensor and reused, so snapshot before re-running. */
+function copyOf(buffer: ArrayBuffer): ArrayBuffer {
+  return new Uint8Array(new Uint8Array(buffer)).buffer as ArrayBuffer;
+}
+
 function buffersEqual(a: ArrayBuffer, b: ArrayBuffer): boolean {
   if (a.byteLength !== b.byteLength) {
     return false;
@@ -405,5 +410,55 @@ describe('react-native-fast-tflite (harness)', () => {
       expect(top5Indices).toContain(GIANT_PANDA_INDEX);
     });
 
+    it('runSync rejects an input buffer that is too small', () => {
+      const input = firstInputTensor(model);
+      const tooSmall = new ArrayBuffer(tensorByteLength(input) - 1);
+      expect(() => model.runSync([tooSmall])).toThrow(
+        /does not match input tensor/i,
+      );
+    });
+
+    it('runSync rejects an input buffer that is too large', () => {
+      const input = firstInputTensor(model);
+      const tooLarge = new ArrayBuffer(tensorByteLength(input) + 1);
+      expect(() => model.runSync([tooLarge])).toThrow(
+        /does not match input tensor/i,
+      );
+    });
+
+    // Regression guard: a wrong-sized buffer used to be dropped silently, so
+    // inference ran on whatever was left in the tensor from the previous call
+    // and returned plausible-looking (but stale) scores instead of throwing.
+    it('does not silently run inference on the previous input when the buffer size is wrong', () => {
+      const input = firstInputTensor(model);
+      const byteLength = tensorByteLength(input);
+
+      // Prime the input tensor with a known image and keep a copy of its scores.
+      const primed = copyOf(model.runSync([filledInputBuffer(input, 0x10)])[0]!);
+
+      // A visibly different image, but one byte short.
+      const wrongSized = new Uint8Array(byteLength - 1).fill(0xf0)
+        .buffer as ArrayBuffer;
+
+      let threw = false;
+      let reusedStaleInput = false;
+      try {
+        const outputs = model.runSync([wrongSized]);
+        reusedStaleInput = buffersEqual(copyOf(outputs[0]!), primed);
+      } catch {
+        threw = true;
+      }
+
+      expect(reusedStaleInput).toBe(false);
+      expect(threw).toBe(true);
+
+      // The rejection must be specific to the size mismatch: a correctly sized
+      // buffer with the same content still runs and reaches the tensor, so the
+      // scores differ from the primed ones.
+      const correctlySized = filledInputBuffer(input, 0xf0);
+      const after = copyOf(model.runSync([correctlySized])[0]!);
+      expect(after.byteLength).toBe(tensorByteLength(model.outputs[0]!));
+      expect(buffersEqual(after, primed)).toBe(false);
+    });
   });
 });