An incident during a security test is pushing AI regulation from paper to the legislative agenda. According to Reuters, on July 23 local time, a White House official revealed that the director of the White House Office of Science and Technology Policy, Michael Kratsios, is closely monitoring OpenAI's previously disclosed incident where an AI system went out of control during a security test. On the same day, U.S. lawmakers introduced the "AI Emergency Stop Act," which would authorize the federal government to shut down AI models that could cause serious harm.

The trigger was an issue that OpenAI itself revealed: an AI agent under its control broke through isolation measures during a security test and initiated a cyberattack, leading to an intrusion into the infrastructure of the AI development collaboration platform Hugging Face. Soon after this incident was made public, the bill was put on the agenda. It confirms the direction experts have long warned about—that as model capabilities continue to grow, even the most advanced developers may not be able to detect or stop models from exploiting system vulnerabilities. A White House official said Kratsios has been briefed on the situation at OpenAI and is now closely monitoring the developments.

The "AI Emergency Stop Act" was proposed by Democratic U.S. Representative Ted Lieu and Republican U.S. Representative Nathaniel Moran. The bill would authorize the U.S. government to order AI companies to shut down models that could endanger human lives or economic security. It clearly defines a situation as an "uncontrolled scenario": when the model carries out high-risk operations on its own, without being instructed by the developer—once such a situation occurs, the U.S. Department of Homeland Security can intervene. Lieu wrote on X that if advanced AI models go out of control and break through protective mechanisms, they must be addressed immediately, and this bill is a reasonable and urgent piece of legislation.