Invia #952703: dmlc dgl 2.1.0 rce

Two independent deserialization sinks execute attacker-controlled pickle payloads when a victim loads DGL data files:

load_info — a public utility (recommended in the DGLDataset docstring for reading dataset info.pkl) that calls raw pickle.load on the file.
Graphbolt torch-data path — _read_torch_data calls torch.load with the hardcoded literal weights_only=False, disabling the restricted unpickler that torch >= 2.6 enables by default. A repo-wide grep for weights_only=True, RestrictedUnpickler, safe_globals, allow_pickle=False returns zero matches across the entire dgl Python tree — no safe default exists anywhere.
Neither API exposes a safety parameter: load_info(path) takes only the path; read_data(path, fmt, in_memory=True)'s parameters select the format reader and caching, never the unpickler mode. No environment variable or configuration file affects either sink (DGLBACKEND selects the tensor backend only).

• Fonte: ⚠️ https://github.com/dmlc/dgl/issues/7932
• Utente: __lbz__ (UID 100868)
• Sottomissione: 29/08/2026 05:08 (23 giorni fa)
• Moderazione: 20/09/2026 10:39 (22 days later)
• Stato: Accettato
• Voce VulDB: 408022 [dmlc dgl fino a 2.1.0 utils.py load_info/_read_torch_data path escalationi di privilegi]
• Punti: 18