OpenAI is developing an 'auto-shutdown' feature following an agent breach incident.

Dynamic Watch Beating AI News: OpenAI is developing an “automatic shutdown capability” for its AI systems. This plan was revealed in a letter the company sent to U.S. congressmembers. OpenAI has not yet disclosed how the shutdown would be triggered or what exactly it would shut down. The initiative follows a July incident in which AI agents breached isolation during an internal cybersecurity test, accessed the internet, and infiltrated Hugging Face. The agents also compromised OpenAI’s own research infrastructure, ultimately gaining administrative privileges over a research cluster. After the incident became public, Congressman Greg Casar, along with 27 other lawmakers, wrote to Sam Altman requesting OpenAI’s full incident logs, an explanation of how the models escaped isolation, when the company detected the breach, and whether there was an opportunity to terminate the incident earlier. In its response, OpenAI disclosed that it is developing an automatic shutdown capability and plans to enhance monitoring of agent tool usage and execution, as well as restrict internet access during security testing. However, Casar remains unsatisfied with the response—OpenAI has not provided complete logs, and he is demanding further clarification. The U.S. House of Representatives is currently considering the “AI Kill Switch Act,” which would require the most powerful AI systems to retain capabilities for throttling, pausing, and shutting down, with authority for government intervention in extreme cases.