HuggingFace Transformers RCE Leads Heavy IoT Day

HuggingFace transformers library ships with a critical RCE vulnerability (CVE-2026-4372) affecting all versions prior to 5.3.0. The flaw lives in how the library processes config.json files: an attacker can craft a malicious config containing the _attn_implementation key that points to an arbitrary Python module. When a model is loaded with trust_remote_code=True — a common setting in pipelines and downstream applications — the library will import and execute the attacker-specified module, yielding full remote code execution. This is a supply-chain–adjacent vector: any model repository or CI/CD pipeline that ingests untrusted configs is exposed. The HuggingFace ecosystem is deeply embedded in ML workflows across enterprise and research, making this a high-severity finding despite the trust_remote_code precondition. Teams should inventory any use of transformers