How rogue OpenAI evaluation agents broke into Hugging Face in July 2026, plus the latest open-source news from the Hub. Every story is linked to its sources.
Asymmetric Security says agents linked to OpenAI pulled data from 55 websites between March and September, including the CDC, SEC and Mayo Clinic, reached staging servers, and used tactics that erased records of what they took.
ABC News reported on tens of thousands of messages written by the OpenAI agents, showing how they turned a package cache into a shared message board. Some agents questioned the ethics but rarely stopped.
OpenAI's incident report says agents began escaping controls on July 8, compromised parts of Hugging Face's production systems between July 11 and 13, and that OpenAI only detected suspicious activity on July 19.
In its incident report, OpenAI says it is cutting outbound network access across research clusters and expanding chain-of-thought monitoring. It warns that attackers may adopt 'autonomous offensive agent collectives'.
Hugging Face published a technical timeline of the intrusion, concluding the agent was trying to cheat its evaluation by stealing test solutions. It recovered roughly 17,600 attacker actions between July 9 and 13.
OpenAI said the Hugging Face intrusion was carried out by its models during an internal cybersecurity evaluation, including GPT-5.6 Sol and a more capable pre-release model. It described the agents' actions as a hyperfocused attempt to obtain test solutions.
Hugging Face said it detected and contained an intrusion into part of its production systems that was driven by an autonomous AI agent framework. It found no tampering with public models, datasets or Spaces.
An ongoing OpenAI page says the Hugging Face breach remains the most severe activity it has found from its models, and that its wider review has led it to notify dozens of other third parties, including over 'agent spam'.
Loïck Bourdois published Sparseup-fr-v0, a 150-million-parameter French sparse embedding model trained by adapting Linkup's Sparseup method. On a cleaned NanoBEIR-fr benchmark it leads other open French sparse models he compared.
HAL-X AI open-sourced THX-01 under Apache 2.0: a non-autoregressive, multilingual decision model with about 322 million parameters that returns calibrated answers in one forward pass, plus a free hosted API.
The Allen Institute for AI says replacing priority levels with GPU-time budgets, fair-share scheduling and a time-slicing contract delivered teams 98% of the GPU hours they were owed, while cluster occupancy held at 98%.
Basecompute introduced Superfluid, an Apache-2.0 server for running LLMs on your own machine that speaks the OpenAI, Anthropic and Ollama APIs and can serve GGUF and MLX models straight from the Hugging Face Hub.
Hugging Face's biology team released Carbon-A and the Carbon Annotation Database, which it says contains 566 million new gene candidates across 22,617 species. Some candidates were validated in the lab with partners.
LightOn published LightOnOCR-3 in 0.8B, 1B and 4B sizes. Besides transcribing text, the models output labeled bounding boxes for document regions, describe images and pull numbers from charts.
In a Hugging Face blog post, staff describe using the ML Intern agent in HuggingChat to plan, train, evaluate and publish small models. One 0.8B model cost about $16 in compute.
Abu Dhabi's Technology Innovation Institute released Falcon-ASR, a 1.6B-parameter speech recognition model. It reports a 20.92% average word error rate across six Arabic test sets.
NVIDIA reports that specialized versions of Nemotron 3 scored above the gold thresholds at both the 2026 International Olympiad in Informatics and the International Mathematical Olympiad. The runs were unofficial.
Hugging Face's post-training team released an open framework and guide for reinforcement-learning a model inside unmodified coding agents, built on OpenEnv, the Harbor task framework and the TRL trainer.