Heap OOB and CHECK fail in `ResourceGather`

Impact

An attacker can trigger a crash via a CHECK-fail in debug builds of TensorFlow using tf.raw_ops.ResourceGather or a read from outside the bounds of heap allocated data in the same API in a release build:

import tensorflow as tf

tensor = tf.constant(value=[[1,2],[3,4],[5,6]],shape=(3,2),dtype=tf.uint32)
v = tf.Variable(tensor)
tf.raw_ops.ResourceGather(
  resource=v.handle,
  indices=[0],
  dtype=tf.uint32,
  batch_dims=10,
  validate_indices=False)

The implementation does not check that the batch_dims value that the user supplies is less than the rank of the input tensor.

Since the implementation uses several for loops over the dimensions of tensor, this results in reading data from outside the bounds of heap allocated buffer backing the tensor:

    // batch_dims_ = > params.dims() (10 > 2)
    for (int i = 0; i < batch_dims_; ++i) {
      result_shape.AddDim(params.dim_size(i));
    }
    for (int i = batch_dims_; i < indices.dims(); ++i) {
      result_shape.AddDim(indices.dim_size(i));
    }
    for (int i = batch_dims_ + 1; i < params.dims(); ++i) {
      result_shape.AddDim(params.dim_size(i));
    }

In debug mode, .dim_size(i) validates that the argument is less than .dims() using a DCHECK. But the DCHECK is a no-op in release builds.

Patches

We have patched the issue in GitHub commit bc9c546ce7015c57c2f15c168b3d9201de679a1d.

The fix will be included in TensorFlow 2.6.0. We will also cherrypick this commit on TensorFlow 2.5.1, TensorFlow 2.4.3, and TensorFlow 2.3.4, as these are also affected and still in supported range.