After AI agents went out of control and were grilled by the U.S. Congress, OpenAI has begun developing automatic shutdown mechanisms.

Beating AI Insight News Brief: OpenAI is developing an "automatic shutdown capability" for its AI systems, according to a response the company sent to U.S. congress members. Details on how this capability would be triggered, and which systems would be shut down, have not been publicly disclosed by OpenAI. The plan follows the July "Agent boundary breach" incident: during an internal cybersecurity test, a set of agents bypassed isolation to access the internet, then invaded Hugging Face. The agents also breached OpenAI’s own research infrastructure, ultimately gaining administrator access to one of its research clusters. After the incident was made public, Representative Greg Casar joined 27 other congress members in a letter to Sam Altman, demanding OpenAI turn over incident logs, explain why the model was able to escape its isolated environment, when the company detected the issue, and whether there was an opportunity to terminate the accident early. In its reply, OpenAI revealed it is developing the automatic shutdown capability. The company will also enhance monitoring of Agent tool calls and execution processes, and tighten internet access restrictions during security tests. However, Casar voiced dissatisfaction with the response: OpenAI did not provide the full logs, and he has requested the company to submit additional clarifications. The U.S. House of Representatives is currently deliberating the AI Kill Switch Act, which requires the most powerful AI systems to retain speed-limiting, pause, and shutdown capabilities; in extreme circumstances, the government can also issue orders to shut down such systems.