TensorFlow is an end-to-end open source platform for machine learning. An attacker can cause a denial of service via a FPE runtime error in `tf.raw_ops.DenseCountSparseOutput`. This is because the implementation(https://github.com/tensorflow/tensorflow/blob/efff014f3b2d8ef6141da30c806faf141297eca1/tensorflow/core/kernels/count_ops.cc#L123-L127) computes a divisor value from user data but does not check that the result is 0 before doing the division. Since `data` is given by the `values` argument, `num_batch_elements` is 0. The fix will be included in TensorFlow 2.5.0. We will also cherrypick this commit on TensorFlow 2.4.2, and TensorFlow 2.3.3, as these are also affected.

In affected versions of TensorFlow running an LSTM/GRU model where the LSTM/GRU layer receives an input with zero-length results in a CHECK failure when using the CUDA backend. This can result in a query-of-death vulnerability, via denial of service, if users can control the input to the layer. This is fixed in versions 1.15.5, 2.0.4, 2.1.3, 2.2.2, 2.3.2, and 2.4.0.

In affected versions of TensorFlow the tf.raw_ops.ImmutableConst operation returns a constant tensor created from a memory mapped file which is assumed immutable. However, if the type of the tensor is not an integral type, the operation crashes the Python interpreter as it tries to write to the memory area. If the file is too small, TensorFlow properly returns an error as the memory area has fewer bytes than what is needed for the tensor it creates. However, as soon as there are enough bytes, the above snippet causes a segmentation fault. This is because the allocator used to return the buffer data is not marked as returning an opaque handle since the needed virtual method is not overridden. This is fixed in versions 1.15.5, 2.0.4, 2.1.3, 2.2.2, 2.3.2, and 2.4.0.

In affected versions of TensorFlow the tf.raw_ops.DataFormatVecPermute API does not validate the src_format and dst_format attributes. The code assumes that these two arguments define a permutation of NHWC. This can result in uninitialized memory accesses, read outside of bounds and even crashes. This is fixed in versions 1.15.5, 2.0.4, 2.1.3, 2.2.2, 2.3.2, and 2.4.0.

In affected versions of TensorFlow under certain cases a saved model can trigger use of uninitialized values during code execution. This is caused by having tensor buffers be filled with the default value of the type but forgetting to default initialize the quantized floating point types in Eigen. This is fixed in versions 1.15.5, 2.0.4, 2.1.3, 2.2.2, 2.3.2, and 2.4.0.

In affected versions of TensorFlow under certain cases, loading a saved model can result in accessing uninitialized memory while building the computation graph. The MakeEdge function creates an edge between one output tensor of the src node (given by output_index) and the input slot of the dst node (given by input_index). This is only possible if the types of the tensors on both sides coincide, so the function begins by obtaining the corresponding DataType values and comparing these for equality. However, there is no check that the indices point to inside of the arrays they index into. Thus, this can result in accessing data out of bounds of the corresponding heap allocated arrays. In most scenarios, this can manifest as unitialized data access, but if the index points far away from the boundaries of the arrays this can be used to leak addresses from the library. This is fixed in versions 1.15.5, 2.0.4, 2.1.3, 2.2.2, 2.3.2, and 2.4.0.

In TensorFlow Lite before versions 1.15.4, 2.0.3, 2.1.2, 2.2.1 and 2.3.1, saved models in the flatbuffer format use a double indexing scheme: a model has a set of subgraphs, each subgraph has a set of operators and each operator has a set of input/output tensors. The flatbuffer format uses indices for the tensors, indexing into an array of tensors that is owned by the subgraph. This results in a pattern of double array indexing when trying to get the data of each tensor. However, some operators can have some tensors be optional. To handle this scenario, the flatbuffer model uses a negative `-1` value as index for these tensors. This results in special casing during validation at model loading time. Unfortunately, this means that the `-1` index is a valid tensor index for any operator, including those that don't expect optional inputs and including for output tensors. Thus, this allows writing and reading from outside the bounds of heap allocated arrays, although only at a specific offset from the start of these arrays. This results in both read and write gadgets, albeit very limited in scope. The issue is patched in several commits (46d5b0852, 00302787b7, e11f5558, cd31fd0ce, 1970c21, and fff2c83), and is released in TensorFlow versions 1.15.4, 2.0.3, 2.1.2, 2.2.1, or 2.3.1. A potential workaround would be to add a custom `Verifier` to the model loading code to ensure that only operators which accept optional inputs use the `-1` special value and only for the tensors that they expect to be optional. Since this allow-list type approach is erro-prone, we advise upgrading to the patched code.

In tensorflow-lite before versions 1.15.4, 2.0.3, 2.1.2, 2.2.1 and 2.3.1, if a TFLite saved model uses the same tensor as both input and output of an operator, then, depending on the operator, we can observe a segmentation fault or just memory corruption. We have patched the issue in d58c96946b and will release patch releases for all versions between 1.15 and 2.3. We recommend users to upgrade to TensorFlow 1.15.4, 2.0.3, 2.1.2, 2.2.1, or 2.3.1.

In tensorflow-lite before versions 1.15.4, 2.0.3, 2.1.2, 2.2.1 and 2.3.1, a crafted TFLite model can force a node to have as input a tensor backed by a `nullptr` buffer. This can be achieved by changing a buffer index in the flatbuffer serialization to convert a read-only tensor to a read-write one. The runtime assumes that these buffers are written to before a possible read, hence they are initialized with `nullptr`. However, by changing the buffer index for a tensor and implicitly converting that tensor to be a read-write one, as there is nothing in the model that writes to it, we get a null pointer dereference. The issue is patched in commit 0b5662bc, and is released in TensorFlow versions 1.15.4, 2.0.3, 2.1.2, 2.2.1, or 2.3.1.

In tensorflow-lite before versions 1.15.4, 2.0.3, 2.1.2, 2.2.1 and 2.3.1, when determining the common dimension size of two tensors, TFLite uses a `DCHECK` which is no-op outside of debug compilation modes. Since the function always returns the dimension of the first tensor, malicious attackers can craft cases where this is larger than that of the second tensor. In turn, this would result in reads/writes outside of bounds since the interpreter will wrongly assume that there is enough data in both tensors. The issue is patched in commit 8ee24e7949a203d234489f9da2c5bf45a7d5157d, and is released in TensorFlow versions 1.15.4, 2.0.3, 2.1.2, 2.2.1, or 2.3.1.

In tensorflow-lite before versions 1.15.4, 2.0.3, 2.1.2, 2.2.1 and 2.3.1, to mimic Python's indexing with negative values, TFLite uses `ResolveAxis` to convert negative values to positive indices. However, the only check that the converted index is now valid is only present in debug builds. If the `DCHECK` does not trigger, then code execution moves ahead with a negative index. This, in turn, results in accessing data out of bounds which results in segfaults and/or data corruption. The issue is patched in commit 2d88f470dea2671b430884260f3626b1fe99830a, and is released in TensorFlow versions 1.15.4, 2.0.3, 2.1.2, 2.2.1, or 2.3.1.

TensorFlow before 1.7.0 has an integer overflow that causes an out-of-bounds read, possibly causing disclosure of the contents of process memory. This occurs in the DecodeBmp feature of the BMP decoder in core/kernels/decode_bmp_op.cc.

In TensorFlow before 1.15.2 and 2.0.1, converting a string (from Python) to a tf.float16 value results in a segmentation fault in eager mode as the format checks for this use case are only in the graph mode. This issue can lead to denial of service in inference/training where a malicious attacker can send a data point which contains a string instead of a tf.float16 value. Similar effects can be obtained by manipulating saved models and checkpoints whereby replacing a scalar tf.float16 value with a scalar string will trigger this issue due to automatic conversions. This can be easily reproduced by tf.constant("hello", tf.float16), if eager execution is enabled. This issue is patched in TensorFlow 1.15.1 and 2.0.1 with this vulnerability patched. TensorFlow 2.1.0 was released after we fixed the issue, thus it is not affected. Users are encouraged to switch to TensorFlow 1.15.1, 2.0.1 or 2.1.0.

In TensorFlow before 1.15, a heap buffer overflow in UnsortedSegmentSum can be produced when the Index template argument is int32. In this case data_size and num_segments fields are truncated from int64 to int32 and can produce negative numbers, resulting in accessing out of bounds heap memory. This is unlikely to be exploitable and was detected and fixed internally in TensorFlow 1.15 and 2.0.

Google TensorFlow 1.7.x and earlier is affected by a Buffer Overflow vulnerability. The type of exploitation is context-dependent.

NULL pointer dereference in Google TensorFlow before 1.12.2 could cause a denial of service via an invalid GIF file.

Memcpy parameter overlap in Google Snappy library 1.1.4, as used in Google TensorFlow before 1.7.1, could result in a crash or read from other parts of process memory.

Invalid memory access and/or a heap buffer overflow in the TensorFlow XLA compiler in Google TensorFlow before 1.7.1 could cause a crash or read from other parts of process memory via a crafted configuration file.

Google TensorFlow 1.7 and below is affected by: Buffer Overflow. The impact is: execute arbitrary code (local).

Google TensorFlow 1.6.x and earlier is affected by: Null Pointer Dereference. The type of exploitation is: context-dependent.