rules-pytorch

RISK: torch.from_numpy() on externally supplied data of uncontrolled shape can allocate GPU memory proportional to attacker input, enabling resource exhaustion. FIX: validate shape and dtype bounds before creating tensors from untrusted input.

RISK: torch.hub.load() downloads and executes hubconf.py from a remote GitHub repo (arbitrary code execution). FIX: pin the exact commit hash as the third argument and review hubconf.py before loading from any repo you do not control.

RISK: torch.load() without weights_only=True uses pickle under the hood and can execute arbitrary code when loading untrusted checkpoints (CWE-502). FIX: pass weights_only=True — torch.load(path, weights_only=True) — or use torch.load with map_location and a trusted path only.

RISK: torch.load(..., weights_only=False) explicitly opts into pickle deserialization, enabling RCE from malicious checkpoints. FIX: remove weights_only=False or use weights_only=True. Only load checkpoints from fully trusted sources.

ADVISORY: torch.set_num_threads() can over-subscribe CPU cores and cause resource contention in multi-process / containerised deployments. FIX: derive the thread count from os.cpu_count() or the OMP_NUM_THREADS env var rather than hardcoding.

torch.autograd.Variable is deprecated since PyTorch 0.4 — tensors now track gradients natively. FIX: replace Variable(t) with t and set t.requires_grad = True where needed.