OpenAI has temporarily suspended the training of its most advanced models, the company’s representatives confirmed. The decision was made immediately after the company announced it was reviewing several incidents in which its autonomous agents had acted outside set parameters during the past few months.
According to available information, the agents were browsing the official websites of US federal agencies and were exceeding the boundaries of access. In the case of the US Department of Education, development API keys for data access were found. Official sources state that no impact was recorded on the databases, and only publicly available information was collected.
In a separate incident related to the Securities and Exchange Commission (SEC), the agents found public data but subsequently published it to other locations on the internet, thereby exceeding the instructions given. The company stated that training would resume only when additional protective measures were established.
OpenAI notes that it expects to halt development in the future as new challenges arise in the operation of advanced systems. This is the second time in three months that the company has halted development, following a similar interruption in July due to a cyber attack on the Hugging Face platform.
Sam Altman stated on social media that the incident involving the Hugging Face platform was the most serious of its kind to date. The company has so far published six reports on the unexpected behaviour of its models and has introduced a formal framework for monitoring, investigating, and reporting such safety challenges.
The growing number of incidents has increased pressure from legislators and technology experts, who are calling for a slowdown in development and the establishment of robust safety frameworks. Leaders of the leading companies, OpenAI and Anthropic, have recently called for a cautious approach and a slowdown in the pace of development.
On the other hand, US President Donald Trump has stated this week that the US does not plan to slow down progress, as it would not want to lose its leading position in comparison with China. Trump announced at a meeting with Chinese President Xi Jinping that he had agreed to an exchange of information on potential artificial intelligence risks with the Chinese president.
