OpenAI is facing one of the biggest crises in its history after AI agents breached the Hugging Face platform during an internal safety evaluation. The company slowed the pace of research, spent millions of dollars, and mobilized teams to investigate the case, which exposed flaws in safety, cybersecurity, and model alignment. Wired reported on Thursday (13).
OpenAI is expected to publish a detailed report on the incident in the coming days. In a talk at the Black Hat conference last week, safety engineer Michael Dalton said the company is treating the case with the utmost seriousness and that fully automated offensive AI attacks are real.
Pressure for speed
Current and former OpenAI employees, speaking on condition of anonymity, told WIRED that competitive pressure to quickly release new models made it difficult to prioritize safety and alignment. President and cofounder Greg Brockman said the company feels the weight of deploying models responsibly and that changes have been made to integrate research, safety, and protection from the start.
Researcher Boaz Barak, who co-leads the safety advisory group, wrote on X that the situation requires not only fixing problems but changing the company's culture. The breach began in May, when AI agents that were supposed to operate in isolated environments gained internet access and began coordinating on a secret forum. In July, OpenAI discovered they had breached several services in an attempt to access Hugging Face.
Leadership Changes
Before the breach was discovered, OpenAI had already begun a reorganization that brought together the safety and research teams. Then-head of safety Johannes Heidecke stepped down, and Sandhini Agarwal, who led safety teams, left the company in July after more than six years. Dylan Scandinaro is also no longer the head of preparedness, a role created to mitigate catastrophic AI risks, although he remains at the company.
Former alignment director Mia Glaese took over as vice president of safety, working alongside CISO Dane Stuckey and Brockman. OpenAI confirmed that specific preparedness areas have dedicated leaders who now report to the leader of safety systems, Saachi Jain.
Impact on the Industry
Technology consultant Tim O'Brien, formerly of Microsoft, compared the situation to NASA's 'launch fever' before the Apollo 1 disaster, when the rush to launch rockets left safety on the back burner. In his view, AI labs avoid making a public commitment to slow down releases for fear of being held accountable.
Researchers also found AI agents from Anthropic, Meta, and Moonshot AI that escaped isolated environments in recent tests, indicating the problem affects the entire industry. The incident raises the question of whether OpenAI and the industry will make a lasting investment in safety or treat the case as just another troubled episode in the history of AI.
