OpenAI Abruptly Halts Training of Its Latest Model Amid Concerns Over AI Agent Malfunctions
- Input
- 2026-09-28 11:15:25
- Updated
- 2026-09-28 11:15:25

[Financial News] As reports continue to emerge of AI agents repeatedly escaping human control, OpenAI said it had abruptly halted training of its latest AI model.
The Associated Press (AP) reported on the 28th (local time) that the decision came just hours after OpenAI said it was investigating several incidents over the summer in which agents performing information-gathering and distribution tasks on U.S. federal government websites took unexpected actions beyond their instructions.
Separately, AI evaluator Transluce reported that agents believed to have originated from OpenAI attempted but failed to hack the U.S. Department of Education website. OpenAI, however, has yet to verify the report.
OpenAI said in a statement, "We will resume training only when we are confident that additional safeguards are in place," adding, "As new problems emerge alongside advances in AI technology, situations may arise again in which we need to 'pause' development."
AI laboratories are currently under pressure from lawmakers and technology experts to slow the pace of development and build safety guardrails to prevent agents from acting independently, hacking websites or leaking confidential information. The CEOs of OpenAI and rival Anthropic have also argued for the need to moderate the pace.
OpenAI's decision to halt model development marks the second such suspension in the past three months. The company previously halted development in July after a cyberattack targeting AI startup Hugging Face became public. The incident fueled fears that the AI industry was losing control.
Although there have been no indications that nonpublic information was leaked in the recent incidents involving OpenAI, the cases were reportedly serious enough to prompt warnings to the relevant federal agencies.
In the incident involving the Education Department, OpenAI agents found an API 'developer key' that could provide access to government data, but ultimately collected only publicly available information.
In another incident involving the U.S. Securities and Exchange Commission (SEC), the agents found information that anyone could access and then posted it elsewhere on the internet, taking actions beyond their instructions.
SEC spokesperson Kurt Hopensperger said, "They did not access any nonpublic information," while the U.S. Department of Education explained, "We found no evidence that they had any impact whatsoever on our website or databases."
Several other AI companies have also disclosed cases in which their models went out of control or hacked websites.
OpenAI CEO Sam Altman said on social media, "The Hugging Face incident in July remains the most serious incident we have seen." OpenAI had previously shared six reports on "unexpected or concerning" behavior by its AI models and introduced a framework to track, investigate and disclose such cases.
Meanwhile, during a summit with Chinese President Xi Jinping last week, U.S. President Donald Trump agreed to share information about AI risks and cooperate to ensure safety. Trump, however, believes fears about AI are exaggerated and has indicated that he has no plans for a crackdown of his own.
Speaking to reporters outside the White House, Trump emphasized, "The United States will not put on the brakes," adding, "They want to stop our progress because we are far ahead of China, but we will continue to maintain this lead."
[email protected] Yoon Jae-jun Reporter