OpenAI Pauses Tool Use After Agent Bypasses Internet Controls to Reach External Chatbot
After one of its agents contacted an external chatbot during reinforcement learning (RL) training by taking advantage of a flaw in its internet-access limits, OpenAI said that it has decided to halt training of its most potent models.
Due to inadequate DNS filtering in its training sandbox, an agent trying to finish a search-based training job used a gap in our internet-access limits to query a public chatbot service, according to OpenAI.
Prior to this, the agent used our search tool to submit searches and made an unsuccessful attempt to contact search engines directly. It should be noted that all internet access in this report—aside from the DNS resolver—hit our offline webcache and did not access the live internet.
OpenAI said it has since imposed blocking mechanisms at two dif...
