diff --git a/torchrl/data/tensor_specs.py b/torchrl/data/tensor_specs.py
index c00167b5b29..49ebdc57f1c 100644
--- a/torchrl/data/tensor_specs.py
+++ b/torchrl/data/tensor_specs.py
@@ -2632,6 +2632,7 @@ def encode(
     ) -> torch.Tensor | TensorDictBase:
         return val
 
+
 class _UnboundedMeta(abc.ABCMeta):
     def __call__(cls, *args, **kwargs):
         instance = super().__call__(*args, **kwargs)
@@ -4930,7 +4931,7 @@ def rand(self, shape: torch.Size = None) -> TensorDictBase:
                 _dict[key] = item.rand(shape)
         # No need to run checks since we know Composite is compliant with
         # TensorDict requirements
-        return TensorDict._new_unsafe(
+        return TensorDict(
             _dict,
             batch_size=_size([*shape, *self.shape]),
             device=self._device,
diff --git a/torchrl/envs/gym_like.py b/torchrl/envs/gym_like.py
index aa1883958ca..bb5a4ddea43 100644
--- a/torchrl/envs/gym_like.py
+++ b/torchrl/envs/gym_like.py
@@ -342,7 +342,7 @@ def _step(self, tensordict: TensorDictBase) -> TensorDictBase:
                     for key, val in TensorDict(obs_dict, []).items(True, True)
                 )
         else:
-            tensordict_out = TensorDict._new_unsafe(
+            tensordict_out = TensorDict(
                 obs_dict,
                 batch_size=tensordict.batch_size,
             )
@@ -376,7 +376,8 @@ def _reset(
 
         source = self.read_obs(obs)
 
-        tensordict_out = TensorDict._new_unsafe(
+        # _new_unsafe cannot be used because it won't wrap non-tensor correctly
+        tensordict_out = TensorDict(
             source=source,
             batch_size=self.batch_size,
         )