From 2ce7dd41d092f6be931619c821ad9a12373876fd Mon Sep 17 00:00:00 2001
From: varun netrakar <varun@security.bugqore.com>
Date: Wed, 14 Oct 2026 13:56:31 +0530
Subject: [PATCH] Fix out-of-bounds write in SetStringData for rank-2 string
 inputs

---
 .../tensorflow/tflite_interpreter_pool.cc     |  9 +++-
 .../tflite_interpreter_pool_test.cc           | 48 +++++++++++++++++++
 2 files changed, 55 insertions(+), 2 deletions(-)

diff --git a/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool.cc b/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool.cc
index 90fed95650a..8278889a2eb 100644
--- a/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool.cc
+++ b/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool.cc
@@ -60,13 +60,17 @@ TfLiteInterpreterWrapper::TfLiteInterpreterWrapper(
 
 absl::Status TfLiteInterpreterWrapper::SetStringData(
     const std::vector<const Tensor*>& tensors, TfLiteTensor* tflite_tensor,
-    int tensor_index, int batch_size) {
+    int tensor_index, int /*batch_size*/) {
   // Format of the buffer for tflite:
   //   [0] number of strings (int32_t)
   //   [4] offset of each string (int32_t)
   //   [sizeof(int32_t) * (num_strings + 1)]] total size of strings
   //   [sizeof(int32_t) * (num_strings + 2)] batch.data()
-  int32_t num_strings = batch_size;
+  // num_strings must be the number of string elements actually written. Sizing
+  // it from the leading (batch) dimension undersizes the offset table whenever
+  // an input has more than one string per batch entry (rank > 1), writing past
+  // the allocation below.
+  int32_t num_strings = 0;
   offset_.clear();
   size_t total_size = 0;
   offset_.push_back(static_cast<int32_t>(total_size));
@@ -75,6 +79,7 @@ absl::Status TfLiteInterpreterWrapper::SetStringData(
     for (int i = 0; i < flat.size(); ++i) {
       total_size += flat(i).size();
       offset_.push_back(static_cast<int32_t>(total_size));
+      ++num_strings;
     }
   }
   size_t required_bytes = total_size + sizeof(int32_t) * (num_strings + 2);
diff --git a/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool_test.cc b/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool_test.cc
index 4df353903c0..bc02a5d2ac7 100644
--- a/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool_test.cc
+++ b/tensorflow_serving/servables/tensorflow/tflite_interpreter_pool_test.cc
@@ -178,6 +178,54 @@ TEST(TfLiteInterpreterWrapper, TfLiteInterpreterWrapperTest) {
       ::testing::ElementsAreArray(expected_strs));
 }
 
+// A string input with rank > 1 has more string elements than its leading
+// dimension. SetStringData must size the tflite string buffer by the number of
+// elements, not the leading dimension, otherwise the offset table is written
+// past the allocation. Empty strings isolate the offset table so the write is a
+// pure out-of-bounds write when the buffer is undersized.
+TEST(TfLiteInterpreterWrapper, SetStringDataRankTwoInput) {
+  std::string model_bytes;
+  TF_ASSERT_OK(ReadFileToString(Env::Default(),
+                                test_util::TestSrcDirPath(kParseExampleModel),
+                                &model_bytes));
+  auto model = tflite::FlatBufferModel::BuildFromModel(
+      flatbuffers::GetRoot<tflite::Model>(model_bytes.data()));
+  tflite::ops::builtin::BuiltinOpResolver resolver;
+  tflite::ops::custom::AddParseExampleOp(&resolver);
+  std::unique_ptr<tflite::Interpreter> interpreter;
+  ASSERT_EQ(tflite::InterpreterBuilder(*model, resolver)(&interpreter,
+                                                         /*num_threads=*/1),
+            kTfLiteOk);
+  ASSERT_EQ(interpreter->inputs().size(), 1);
+  const int idx = interpreter->inputs()[0];
+  auto* tensor = interpreter->tensor(idx);
+  ASSERT_EQ(tensor->type, kTfLiteString);
+  interpreter->ResizeInputTensor(idx, {1});
+  ASSERT_EQ(interpreter->AllocateTensors(), kTfLiteOk);
+
+  auto interpreter_wrapper =
+      std::make_unique<TfLiteInterpreterWrapper>(std::move(interpreter));
+
+  // Shape [1, num_cols]: leading dimension is 1 but there are num_cols strings.
+  const int num_cols = 32;
+  TensorShape shape;
+  shape.AddDim(1);
+  shape.AddDim(num_cols);
+  Tensor t(DT_STRING, shape);
+  for (int i = 0; i < num_cols; ++i) {
+    t.flat<tstring>()(i) = "";
+  }
+  std::vector<const Tensor*> data;
+  data.push_back(&t);
+
+  // The caller derives batch_size from the leading dimension (dim_size(0) == 1).
+  TF_ASSERT_OK(interpreter_wrapper->SetStringData(data, tensor, idx,
+                                                  /*batch_size=*/1));
+  auto wrapped = interpreter_wrapper->Get();
+  auto* tflite_input_tensor = wrapped->tensor(idx);
+  EXPECT_EQ(tflite::GetStringCount(tflite_input_tensor), num_cols);
+}
+
 }  // namespace internal
 }  // namespace serving
 }  // namespace tensorflow