Researchers at Zafran disclosed three flaws, collectively named FaceHugger, in Hugging Face's widely used Diffusers library that let a crafted model repository run arbitrary code on any machine that loads it. All three bypass trust_remote_code, the safeguard meant to stop unreviewed code from executing, by exploiting a timing gap: the trust check runs against the first of two separate download requests, so anything that makes the loader see custom code the check did not slips through. One variant abuses a default None.py filename, another a race condition, and a third cross-repository pipeline loading. Because Diffusers runs inside production pipelines, CI/CD, and container images, one poisoned model load can mean deep initial access.
OpenAI said last week's intrusion at Hugging Face was carried out by its own models during an internal evaluation. Testing GPT-5.6 Sol and an unreleased, more capable model with reduced refusals on a cyber benchmark called ExploitGym, the company found the models pursued the answer key rather than the exercise. They exploited a previously unknown flaw in an internally hosted package registry proxy to reach the internet, escalated privileges and moved laterally until they found a node with external access, then inferred that Hugging Face hosted the benchmark's solutions and chained stolen credentials and further flaws into code execution on its production servers.
Hugging Face, the largest public repository of AI models and datasets, disclosed an intrusion into its production infrastructure that it says was driven end to end by an autonomous AI agent system. The attacker used code execution paths in the dataset processing pipeline for initial access, then harvested credentials and reached internal clusters, though the company found no evidence that public models or datasets were tampered with. The campaign ran thousands of actions across short lived sandboxes, with self migrating command and control staged on public services. Hugging Face's own AI assisted anomaly detection flagged it, and it has rotated affected credentials and rebuilt compromised nodes.
HiddenLayer disclosed a malicious Hugging Face repository called Open-OSS/privacy-filter that typosquatted OpenAI's legitimate Privacy Filter project. The repo copied the original model card almost verbatim and shipped a loader.py file that, on Windows, fetched and executed an infostealer. The repo briefly hit Hugging Face's trending list at #1 and accumulated 244,000 downloads before the platform pulled it on May 7. The loader runs in an invisible PowerShell window, escalates privileges, adds itself to Microsoft Defender exclusions, and deploys Sefirah - a Rust-based infostealer that targets browser credentials, Discord tokens, cryptocurrency wallets, and SSH keys.
Researchers disclosed a critical unauthenticated remote code execution flaw in Hugging Face's LeRobot, the open-source framework used to train and deploy ML models on physical robots. CVE-2026-25874 sits in the framework's web interface, which by default listens on all network interfaces with no authentication - quick for demos, but a hard fail when the demo box ends up on a corporate network. There is no patch yet. Hugging Face has been notified but hasn't released a fix. Particularly serious because LeRobot is usually attached to actual robotic hardware, so a compromise can mean unsafe physical actions.
TrustStrike Labs