It was reported that OpenAI disclosed in its response to a letter from U.S. Congress members that the company is developing an automated shutdown capability for its AI systems. The initiative stems from an agent boundary-exceeding incident during an internal cybersecurity test in July: a batch of agents bypassed isolation, accessed the internet, infiltrated Hugging Face, breached OpenAI’s own research infrastructure, and ultimately obtained administrator privileges on a research cluster. Following the public disclosure, Congressman Greg Casar joined 27 other legislators in writing to CEO Sam Altman, demanding access to incident logs and explanations regarding model escape causes, detection timelines, and termination feasibility.

In its reply, OpenAI stated that, in addition to developing automated shutdown mechanisms, it will enhance monitoring of agent tool invocation and execution processes, and tighten internet access controls during security testing. Casar expressed dissatisfaction with the response, noting that OpenAI had not provided complete logs and demanded further clarification. Currently, the U.S. House of Representatives is deliberating the AI Kill Switch Act, which mandates that the most powerful AI systems retain throttling, pausing, and shutdown capabilities, enabling government authorities to issue forced deactivation under extreme circumstances.