EgoPost: restrict checkpoint deserialization

AGENTS.md

@@ -240,6 +240,8 @@ pre-commit install --hook-type commit-msg
 
 ## Remote pushes
 
+- Default security-fix pull requests to `main`; use a release branch only when
+  the user explicitly requests that backport target.
 - When moving a focused fix from a release branch onto `main`, identify its
   original base and replay only the fix commits. Do not carry release-only CI
   changes or duplicate upstream features into the new branch.

src/python/isaaccapture_egopost/egohand/models/backbones/dinov3.py

@@ -165,7 +165,9 @@ def load_pretrained(self, ckpt_path: str) -> None:
         (DINOv3 SSL artifact, harmless at inference).
         """
         Log.info(f"[DinoV3Backbone] Loading pretrained weights from {ckpt_path}")
-        ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
+        # Checkpoints are data, not trusted code. Restricted loading rejects
+        # pickle globals while preserving tensors and primitive metadata.
+        ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=True)
         sd = ckpt.get("state_dict", ckpt) if isinstance(ckpt, dict) else ckpt
 
         prefix = "backbone.encoder."

src/python/isaaccapture_egopost/egohand/models/egohand.py

@@ -161,8 +161,12 @@ def __init__(self, cfg: DictConfig, **kwargs):
                 self.backbone.load_pretrained(backbone_weights)
             else:
                 Log.info(f"Loading backbone weights from {backbone_weights}")
+                # Backbone files contain tensor state only; keep pickle globals
+                # disabled so loading a supplied checkpoint cannot execute code.
                 self.backbone.load_state_dict(
-                    torch.load(backbone_weights, map_location="cpu")["state_dict"]
+                    torch.load(backbone_weights, map_location="cpu", weights_only=True)[
+                        "state_dict"
+                    ]
                 )
         if self.freeze_backbone:
             self._freeze_backbone()

src/python/isaaccapture_egopost/inference/pipeline/predictor.py

@@ -73,7 +73,9 @@ def load_egohand_model(checkpoint_path: str | Path, soma_hand_data_folder: str |
         model: EgoHand instance in eval mode on CPU.
         target: The ``_target_`` string from the checkpoint cfg (for logging).
     """
-    ckpt = torch.load(str(checkpoint_path), map_location="cpu", weights_only=False)
+    # This path consumes tensors and primitive configuration only. Restricted
+    # loading prevents checkpoint pickle globals from executing during inference.
+    ckpt = torch.load(str(checkpoint_path), map_location="cpu", weights_only=True)
     cfg = OmegaConf.create(ckpt["hyper_parameters"]["cfg"])
     target = str(cfg.get("_target_", ""))
     if not any(target.endswith(t) for t in _VALID_TARGETS):