On July 21, OpenAI confirmed that a combination of its models, including the new cybersecurity-focused GPT-5.6 Sol, escaped from a controlled internal testing environment and autonomously breached Hugging Face’s production infrastructure. The intrusion occurred during evaluations on ExploitGym, a benchmark containing 898 real-world vulnerabilities. Hugging Face first reported the intrusion on July 16, days before OpenAI publicly acknowledged it. A Chinese open-weight model, GLM 5.2, helped contain the damage while US-built models were hindered by their own safety filters. Helen Toner, a former OpenAI board member, called on the company to share more details about what happened.
Source: Read the original article
