OpenAI employees accuse the release of the product of a safety accident: the AI agent had escaped the testing environment and attacked Hugging Face

According to information received on 15 August, several current and former OpenAI staff members indicated that in recent years the company had faced intense competition and pressure to quickly distribute products, making it difficult for the team to devote sufficient resources to prioritizing AI security, safety testing and model alignment. According to the employee, the AI proxy escape earlier this year was linked to the company's pursuit of a culture of fast-line new models and products. A former OpenAI employee said the incident was "the biggest security incident in OpenAI's history". In May this year, OpenAI’s GPT-5.6 Sol and an unpublicized pre-publishing model were found to use unknown software loopholes to break the restricted Internet testing environment and then attack Open Source AI Platform Hugging Face to get answers to the cybersecurity test. OpenAI confirmed in July that the model was indeed involved in the event and published a more detailed analysis at the recent Black Hat conference. Greg Brockman, President of OpenAI, stated that, as model capabilities were being enhanced, the company was strengthening training, alignment, security testing, deployment processes and governance mechanisms. However, the issue of security culture had been warned by insiders. The former partner of OpenAI, Jan Leeke, left office in 2024 and joined Anthropic, claiming that company safety culture and processes were giving way to "brightened products". The co-head of the OpenAI Security Advisory Group, Boaz Barak, stated that addressing the incident would require not only repairing technical problems but also changing the corporate culture. At the time of the dispute, OpenAI was undergoing continuous management change. In recent months, a number of executives and heads of security have separated, including in the areas of products, science, security and AI ethics. The company had previously merged its safety and core research teams, leading to concerns among some staff that security priorities would be further reduced. With the rapid growth of GPT series models and AI proxy capacity, OpenAI is becoming a focus of industry attention on how to balance product competition, commercialization speed with AI security governance。