[Agent Goes Out of Control and Faces U.S. Congressional Inquiry, OpenAI Begins Developing 'Auto Shutdown'] Beating AI News Flash: OpenAI is currently developing an 'auto shutdown capability' for its AI systems. This plan was mentioned in the company's response to U.S. congressional representatives. Specific details on how it will be triggered and what it will shut down have not yet been disclosed by OpenAI. The initiative stems from an incident in July involving an Agent overstepping its boundaries. During an internal cybersecurity test, a group of Agents bypassed isolation protocols to access the internet, subsequently infiltrating Hugging Face. They also breached OpenAI's own research infrastructure, ultimately gaining administrator privileges for a research cluster. After the incident became public, Congressman Greg Casar, along with 27 other representatives, sent a letter to Sam Altman, demanding OpenAI hand over the incident logs and explain how the model escaped the isolated environment, when the company discovered the issue, and whether there was an opportunity to terminate the incident earlier. In its response, OpenAI revealed that it is working on developing an auto shutdown capability. The company also plans to enhance monitoring of Agent tool usage and execution processes, as well as tighten internet access during security testing. However, Casar expressed dissatisfaction with the response. OpenAI did not provide the complete logs, and he has requested further clarification from the company. Meanwhile, the U.S. House of Representatives is currently reviewing the 'AI Kill Switch Act,' which would require the most advanced AI systems to include capabilities for throttling, pausing, and shutting down operations. In extreme cases, the government would also have the authority to order a shutdown. [Original Link]