Hugging Face peft's LoRA-GA and CorDA initialization modules (src/peft/tuners/lora/corda.py lines ~102 and ~163, and src/peft/tuners/lora/loraga.py line ~101) c…

CVE-2026-71281 affects Hugging Face's PEFT library, specifically the LoRA-GA and CorDA initialization modules. The vulnerability stems from insecure use of torch.load() without the weights_only=True parameter in corda.py (lines ~102 and ~163) and loraga.py (line ~101). This bypasses PEFT's own safe-loading wrapper used elsewhere in the codebase. Because torch.load() without weights_only=True performs full pickle deserialization, an attacker can craft a malicious cache or covariance file to achieve arbitrary code execution. The attack vector involves sharing or distributing poisoned LoRA-GA or CorDA cache files that victims may download and load. This is a supply chain risk for ML practitioners using shared model initialization caches. The vulnerability is considered high severity due to the potential for arbitrary code execution.