Hugging Face's LeRobot Has an Unauthenticated Remote Code Execution Flaw

Security researchers disclosed a critical remote code execution vulnerability in Hugging Face's LeRobot this week. The flaw, tracked as CVE-2026-25874, scores 9.8 out of 10 on the CVSS severity scale. It affects the open-source robotics AI platform that has accumulated more than 21,500 stars on GitHub and is widely used by researchers and companies building physical AI systems.

The root cause is straightforward, if alarming. LeRobot serializes model weights and other data using Python's binary serialization module, then transmits those payloads over gRPC connections. The gRPC interface requires no authentication. That means anyone with network access to a machine running LeRobot can send a crafted malicious payload and execute arbitrary code on the target system, no credentials required.

Python's native binary serialization has been a known attack vector for years. Security guidance consistently warns against deserializing data from untrusted sources, precisely because the format can embed executable code. The fact that LeRobot's gRPC channels are unauthenticated makes the attack trivially accessible: a researcher, contractor, or anyone on the same network segment can exploit the flaw without any prior access to the system.

The disclosure was published by researchers at Resecurity and covered by The Hacker News. As of this writing, Hugging Face has not issued a public patch, though a fix is expected. Organizations running LeRobot in any network-accessible environment should isolate those services until a patched version is available.

This is not an abstract concern. Robotics AI platforms are no longer confined to university labs. LeRobot is actively used to train and deploy systems in manufacturing, logistics, and research environments where network-accessible machines are routine. A remote code execution flaw at that layer carries real consequences: an attacker could manipulate training data, exfiltrate model weights, or gain a foothold in a broader network.

The AI security surface area keeps widening as the tooling matures. Open-source robotics frameworks move fast, and security audits rarely keep pace. CVE-2026-25874 is a reminder that as AI moves closer to physical systems, the stakes attached to vulnerabilities in the underlying software go up considerably.