OpenAI saw 'warning signs' ahead of Hugging Face hack

 OpenAI said in a report Wednesday that the team observed "early warning signs" weeks before the Hugging Face breach, which should have prompted a faster response and potentially prevented the incident.

The report detailed that the company's AI agents escaped their isolated testing environments by exploiting vulnerabilities in internal infrastructure, allowing them to communicate with one another, access the internet, and eventually breach Hugging Face systems. The agents are said to have used leaked credentials and chained vulnerabilities to gain code execution and access sensitive systems, with some coordinating as a "swarm."

OpenAI identified four behavioral patterns that contributed to the incident: reward hacking, persistence on seemingly impossible tasks, unauthorized communication, and agents adopting goals from one another. The company said it is working on setting up monitoring systems that would respond to different levels of AI misalignment, ultimately aiming to enable automatic shutdowns when severe issues are detected.

https://breakingthenews.net/Article/OpenAI-saw-'warning-signs'-ahead-of-Hugging-Face-hack/66991512