High severityNVD Advisory· Published Jan 3, 2024· Updated Jun 17, 2025

Stack overflow in paddle.searchsorted

CVE-2023-52304

Description

Stack overflow in paddle.searchsorted in PaddlePaddle before 2.6.0. This flaw can lead to a denial of service, or even more damage.

AI Insight

LLM-synthesized narrative grounded in this CVE's description and references.

Stack overflow in PaddlePaddle's paddle.searchsorted function before version 2.6.0 can lead to denial of service.

Root

Cause

CVE-2023-52304 describes a stack overflow vulnerability in the paddle.searchsorted function of PaddlePaddle, a deep learning framework. The issue exists in versions before 2.6.0 and arises from insufficient input validation. Specifically, the SearchsortedInferMeta function did not enforce that the input sorted_sequence tensor has at least one dimension, and the CastPyArg2Ints (and similar) functions failed to check whether a Python object is a Paddle tensor before treating it as a sequence [1][2]. This allowed a crafted input with zero or malformed dimensions to trigger unbounded recursion or excessive stack allocation.

Attack

Vector

An attacker can supply a specially crafted tensor to paddle.searchsorted that bypasses the missing dimension check. The vulnerability is exploitable without authentication if the attacker can invoke the API (e.g., through a model serving endpoint). The lack of a PADDLE_ENFORCE_GE guard on the sequence dimension size, combined with improper type-checking before sequence iteration, allows the function to enter an infinite or deeply nested loop, exhausting stack memory [2].

Impact

Successful exploitation leads to a stack overflow, causing a denial of service (DoS) by crashing the application. The advisory states that the flaw "can lead to a denial of service, or even more damage," implying potential for arbitrary code execution under certain conditions, though this is not elaborated in the available sources [1]. The vulnerability affects all deployments of PaddlePaddle using the affected function until patched.

Mitigation

The vulnerability is fixed in PaddlePaddle version 2.6.0. The fix adds a dimension check (PADDLE_ENFORCE_GE) in SearchsortedInferMeta and includes a type check (!PyObject_TypeCheck(obj, p_tensor_type)) in the sequence-casting helpers to avoid misidentifying tensors as generic sequences [2]. Users should upgrade to 2.6.0 or later. No workaround is documented. The PyPI advisory database lists this CVE with high severity, urging immediate update [4].

References

AI Insight generated on May 20, 2026. Synthesized from this CVE's description and the cited reference URLs; citations are validated against the source bundle.

Affected packages

Versions sourced from the GitHub Security Advisory.

Package	Affected versions	Patched versions
PaddlePaddlePyPI	< 2.6.0	2.6.0

Affected products

ghsa-coords
pkg:pypi/paddlepaddle
Range: < 2.6.0
PaddlePaddle/PaddlePaddlev5
Range: 0

Patches

19da5c0c4d8c

fix security bug (#55782)

https://github.com/PaddlePaddle/PaddlewanghuancoderAug 2, 2023via ghsa

commit

18 files changed · +90 −6

paddle/fluid/pybind/op_function_common.cc+4 −4 modified

@@ -412,7 +412,7 @@ std::vector<int> CastPyArg2Ints(PyObject* obj,
             i));
       }
     }
-  } else if (PySequence_Check(obj)) {
+  } else if (PySequence_Check(obj) && !PyObject_TypeCheck(obj, p_tensor_type)) {
     Py_ssize_t len = PySequence_Size(obj);
     value.reserve(len);
     PyObject* item = nullptr;
@@ -488,7 +488,7 @@ std::vector<int64_t> CastPyArg2Longs(PyObject* obj,
             i));
       }
     }
-  } else if (PySequence_Check(obj)) {
+  } else if (PySequence_Check(obj) && !PyObject_TypeCheck(obj, p_tensor_type)) {
     Py_ssize_t len = PySequence_Size(obj);
     PyObject* item = nullptr;
     for (Py_ssize_t i = 0; i < len; i++) {
@@ -567,7 +567,7 @@ std::vector<float> CastPyArg2Floats(PyObject* obj,
             i));
       }
     }
-  } else if (PySequence_Check(obj)) {
+  } else if (PySequence_Check(obj) && !PyObject_TypeCheck(obj, p_tensor_type)) {
     Py_ssize_t len = PySequence_Size(obj);
     PyObject* item = nullptr;
     for (Py_ssize_t i = 0; i < len; i++) {
@@ -642,7 +642,7 @@ std::vector<double> CastPyArg2Float64s(PyObject* obj,
             i));
       }
     }
-  } else if (PySequence_Check(obj)) {
+  } else if (PySequence_Check(obj) && !PyObject_TypeCheck(obj, p_tensor_type)) {
     Py_ssize_t len = PySequence_Size(obj);
     PyObject* item = nullptr;
     for (Py_ssize_t i = 0; i < len; i++) {

paddle/phi/infermeta/binary.cc+6 −0 modified

@@ -2663,6 +2663,12 @@ void SearchsortedInferMeta(const MetaTensor& sorted_sequence,
                            MetaTensor* out) {
   auto sequences_dims = sorted_sequence.dims();
   auto values_dims = value.dims();
+  PADDLE_ENFORCE_GE(
+      sequences_dims.size(),
+      1,
+      phi::errors::InvalidArgument(
+          "Input sequences's dimension(%d) must be greater or equal than 1",
+          sequences_dims.size()));
 
   bool flag = true;
   if (sequences_dims.size() != values_dims.size()) {

paddle/phi/kernels/cpu/broadcast_kernel.cc+5 −0 modified

@@ -28,6 +28,11 @@ void BroadcastKernel(const Context& dev_ctx,
                      const DenseTensor& x,
                      int root,
                      DenseTensor* out) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      phi::errors::InvalidArgument("Tensor need be broadcast must not empyt."));
+
 #if defined(PADDLE_WITH_GLOO)
   dev_ctx.template Alloc<T>(out);
   auto comm_context =

paddle/phi/kernels/cpu/dot_kernel.cc+3 −0 modified

@@ -27,6 +27,9 @@ void DotKernel(const Context& dev_ctx,
                const DenseTensor& x,
                const DenseTensor& y,
                DenseTensor* out) {
+  if (out->numel() <= 0) {
+    return;
+  }
   auto const *x_ptr = x.data<T>(), *x_ptr_ = &x_ptr[0];
   auto const *y_ptr = y.data<T>(), *y_ptr_ = &y_ptr[0];
   T* z = dev_ctx.template Alloc<T>(out);

paddle/phi/kernels/cpu/eig_kernel.cc+4 −0 modified

@@ -24,6 +24,10 @@ void EigKernel(const Context& dev_ctx,
                const DenseTensor& x,
                DenseTensor* out_w,
                DenseTensor* out_v) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      errors::InvalidArgument("EigKernel input tensor is empty."));
   if (!IsComplexType(x.dtype())) {
     dev_ctx.template Alloc<phi::dtype::Complex<T>>(out_w);
     dev_ctx.template Alloc<phi::dtype::Complex<T>>(out_v);

paddle/phi/kernels/cpu/reduce_kernel.cc+4 −0 modified

@@ -29,6 +29,10 @@ void ReduceKernel(const Context& dev_ctx,
                   int root,
                   int reduce_type,
                   DenseTensor* out) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      phi::errors::InvalidArgument("Tensor need be reduced must not empyt."));
 #if defined(PADDLE_WITH_GLOO)
   out->Resize(x.dims());
   dev_ctx.template Alloc<T>(out);

paddle/phi/kernels/cpu/top_k_kernel.cc+6 −0 modified

@@ -153,6 +153,12 @@ void TopkKernel(const Context& dev_ctx,
   }
 
   int k = k_scalar.to<int>();
+  PADDLE_ENFORCE_GE(
+      x.numel(),
+      k,
+      errors::InvalidArgument(
+          "x has only %d element, can not find %d top values.", x.numel(), k));
+
   if (k_scalar.FromTensor()) {
     auto out_dims = out->dims();
     // accroding to axis to set K value in the dim

paddle/phi/kernels/funcs/gather_scatter_functor.cc+0 −1 modified

@@ -122,7 +122,6 @@ struct cpu_gather_scatter_functor {
 
           self_idx = is_scatter_like ? replace_index : index_idx;
           src_idx = is_scatter_like ? index_idx : replace_index;
-
           reduce_op((tensor_t*)(self_data + self_idx),  // NOLINT
                     (tensor_t*)(src_data + src_idx));   // NOLINT
           index_idx++;

paddle/phi/kernels/funcs/reduce_function.h+9 −0 modified

@@ -988,6 +988,10 @@ void ReduceKernel(const KPDevice& dev_ctx,
                   const TransformOp& transform,
                   const std::vector<int>& origin_reduce_dims,
                   bool is_mean = false) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      phi::errors::InvalidArgument("Tensor need be reduced must not empyt."));
 #ifdef PADDLE_WITH_XPU_KP
   auto stream = dev_ctx.x_context()->xpu_stream;
 #else
@@ -1298,6 +1302,11 @@ void ReduceKernelImpl(const Context& dev_ctx,
                       const std::vector<int64_t>& dims,
                       bool keep_dim,
                       bool reduce_all) {
+  PADDLE_ENFORCE_GT(
+      input.numel(),
+      0,
+      phi::errors::InvalidArgument("Tensor need be reduced must not empyt."));
+
   dev_ctx.template Alloc<OutT>(output);
 
   if (reduce_all) {

paddle/phi/kernels/funcs/repeat_tensor2index_tensor.h+5 −0 modified

@@ -32,6 +32,11 @@ void RepeatsTensor2IndexTensor(const Context& ctx,
 
   int64_t index_size = 0;
   for (int i = 0; i < repeats.dims()[0]; i++) {
+    PADDLE_ENFORCE_GE(repeats_data[i],
+                      0,
+                      phi::errors::InvalidArgument(
+                          "repeats must grater or equal than 0, but got %d",
+                          repeats_data[i]));
     index_size += repeats_data[i];
   }
   std::vector<RepeatsT> index_vec(index_size);

paddle/phi/kernels/gpu/broadcast_kernel.cu+5 −0 modified

@@ -28,6 +28,11 @@ void BroadcastKernel(const Context& dev_ctx,
                      const DenseTensor& x,
                      int root,
                      DenseTensor* out) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      phi::errors::InvalidArgument("Tensor need be broadcast must not empyt."));
+
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
   dev_ctx.template Alloc<T>(out);
   gpuStream_t stream = dev_ctx.stream();

paddle/phi/kernels/gpu/dot_kernel.cu+3 −0 modified

@@ -31,6 +31,9 @@ void DotKernel(const Context& dev_ctx,
                const DenseTensor& x,
                const DenseTensor& y,
                DenseTensor* out) {
+  if (out->numel() <= 0) {
+    return;
+  }
   dev_ctx.template Alloc<T>(out);
   if (out->dims().size() == 0) {
     auto eigen_out = phi::EigenScalar<T>::From(*out);

paddle/phi/kernels/gpu/lerp_kernel.cu+10 −0 modified

@@ -51,6 +51,16 @@ void LerpKernel(const Context &ctx,
                 const DenseTensor &y,
                 const DenseTensor &weight,
                 DenseTensor *out) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      phi::errors::InvalidArgument("LerpKernel's input x must not empyt."));
+
+  PADDLE_ENFORCE_GT(
+      y.numel(),
+      0,
+      phi::errors::InvalidArgument("LerpKernel's input y must not empyt."));
+
   int rank = out->dims().size();
   PADDLE_ENFORCE_GE(
       rank,

paddle/phi/kernels/gpu/reduce_kernel.cu+4 −0 modified

@@ -29,6 +29,10 @@ void ReduceKernel(const Context& dev_ctx,
                   int root,
                   int reduce_type,
                   DenseTensor* out) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      phi::errors::InvalidArgument("Tensor need be reduced must not empyt."));
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
   out->Resize(x.dims());
   dev_ctx.template Alloc<T>(out);

paddle/phi/kernels/gpu/top_k_kernel.cu+5 −0 modified

@@ -77,6 +77,11 @@ void TopkKernel(const Context& dev_ctx,
   if (axis < 0) axis += in_dims.size();
 
   int k = k_scalar.to<int>();
+  PADDLE_ENFORCE_GE(
+      x.numel(),
+      k,
+      errors::InvalidArgument(
+          "x has only %d element, can not find %d top values.", x.numel(), k));
   if (k_scalar.FromTensor()) {
     phi::DDim out_dims = out->dims();
     out_dims[axis] = k;

paddle/phi/kernels/impl/lerp_kernel_impl.h+10 −0 modified

@@ -83,6 +83,16 @@ void LerpKernel(const Context& ctx,
                 const DenseTensor& y,
                 const DenseTensor& weight,
                 DenseTensor* out) {
+  PADDLE_ENFORCE_GT(
+      x.numel(),
+      0,
+      phi::errors::InvalidArgument("LerpKernel's input x must not empyt."));
+
+  PADDLE_ENFORCE_GT(
+      y.numel(),
+      0,
+      phi::errors::InvalidArgument("LerpKernel's input y must not empyt."));
+
   int rank = out->dims().size();
   PADDLE_ENFORCE_GE(
       rank,

paddle/phi/kernels/impl/repeat_interleave_kernel_impl.h+5 −0 modified

@@ -58,6 +58,11 @@ void RepeatInterleaveKernel(const Context& ctx,
                             int repeats,
                             int dim,
                             DenseTensor* out) {
+  PADDLE_ENFORCE_GT(repeats,
+                    0,
+                    phi::errors::InvalidArgument(
+                        "repeats must grater than 0, but got %d", repeats));
+
   auto place = ctx.GetPlace();
   auto cpu_place = phi::CPUPlace();

python/paddle/tensor/manipulation.py+2 −1 modified

@@ -543,6 +543,8 @@ def unstack(x, axis=0, num=None):
         raise ValueError(
             '`axis` must be in the range [-{0}, {0})'.format(x.ndim)
         )
+    if num is not None and (num < 0 or num > x.shape[axis]):
+        raise ValueError(f'`num` must be in the range [0, {x.shape[axis]})')
     if in_dynamic_mode():
         if num is None:
             num = x.shape[axis]
@@ -4372,7 +4374,6 @@ def repeat_interleave(x, repeats, axis=None, name=None):
     if axis is None:
         x = paddle.flatten(x)
         axis = 0
-
     if in_dynamic_mode():
         if isinstance(repeats, Variable):
             return _C_ops.repeat_interleave_with_tensor_index(x, repeats, axis)

Vulnerability mechanics

Generated on May 9, 2026. Inputs: CWE entries + fix-commit diffs from this CVE's patches. Citations validated against bundle.

References

News mentions

No linked articles in our index yet.

cvss	0.455
epss	0.000
exploit	0.000
kev	0.000
patch	-0.070
ransomware	0.000