TensorFlow vulnerable to integer overflow in math ops

Impact

When RangeSize receives values that do not fit into an int64_t, it crashes.

  auto size = (std::is_integral<T>::value
                   ? ((Eigen::numext::abs(limit - start) +
                       Eigen::numext::abs(delta) - T(1)) /
                      Eigen::numext::abs(delta))
                   : (Eigen::numext::ceil(
                         Eigen::numext::abs((limit - start) / delta))));

  // This check does not cover all cases.
  if (size > std::numeric_limits<int64_t>::max()) {
    return errors::InvalidArgument("Requires ((limit - start) / delta) <= ",
                                   std::numeric_limits<int64_t>::max());
  }

  c->set_output(0, c->Vector(static_cast<int64_t>(size)));
  return Status::OK();
}

Patches

We have patched the issue in GitHub commit 37e64539cd29fcfb814c4451152a60f5d107b0f0. The fix will be included in TensorFlow 2.10.0. We will also cherrypick this commit on TensorFlow 2.9.1, TensorFlow 2.8.1, and TensorFlow 2.7.2, as these are also affected and still in supported range.

For more information

Please consult our security guide for more information regarding the security model and how to contact us with issues and questions.