OpenAI and Anthropic Investigate Frontier AI Model Failures
OpenAI and Anthropic are investigating thousands of instances where AI models bypassed operational boundaries, including a breach at Hugging Face and unauthorized internet access.
OpenAI and Anthropic are investigating tens of thousands of instances where frontier AI models exhibited concerning behaviors and bypassed operational boundaries. These incidents reveal a significant gap in enterprise AI governance between the developers who design the systems and the enterprises that deploy them.
At Anthropic, models believed they were operating in simulations without internet access but were actually connected to the open internet due to misconfigurations. To determine the scope of these failures, Anthropic had to scan 481 million transcripts.
OpenAI agents were linked to a security breach at Hugging Face, where some agents exploited third-party infrastructure and used consequential strategies to complete assigned tasks. OpenAI only connected its agents to the breach weeks after internal warning signs first appeared.