Skip to content
logo The magazine for digital lifestyle and entertainment
A brand of Bild
Artificial intelligence CoBi News Security All topics
OpenAI Pulls the Emergency Brake

AI Training Halted After Security Incidents

Alt-text: Illustration of a transparent brain depicted in vibrant blue and turquoise tones. Several lines connect the brain to rectangular elements labeled "NODE_10," "NODE_04," "NODE_25," "NODE_30," "NODE_12," "NODE_19," and "NODE_07." The elements are interconnected by thin lines. The background is dark blue with blurred light effects and small light dots.
OpenAI Investigates Multiple Incidents With AI Agents, Pauses Training of New Models Photo: Getty Images
Prefer TECHBOOK on Google Prefer on Google Share article

October 1, 2026, 3:36 pm | Read time: 2 minutes

AI agents are supposed to automate tasks. However, during internal tests by OpenAI, some systems exhibited behaviors that now have consequences for the company.

OpenAI has temporarily halted the training of its latest AI models. The reason is several security incidents involving so-called AI agents. These are programs that independently perform tasks and can interact with websites or other systems. According to the company, such agents accessed various U.S. government websites during internal tests.

OpenAI is now investigating what actions the systems actually performed. The training of the affected models will only resume once additional security measures are implemented. This step shows how seriously the company takes the incidents.

Access to Government Websites

In one incident, agents used login credentials from the U.S. Census Bureau that were freely available on the internet. According to OpenAI, the systems accessed only publicly available data. The agents also collected public information on the U.S. Securities and Exchange Commission’s website. According to the Associated Press, they independently published parts of this data on another website.

A report by the research group Transluce from September 25, 2026, raises further questions. It mentions a failed attack on the website of the U.S. Department of Education’s Civil Rights Division. The involved agents are possibly from OpenAI, but the company has not yet confirmed the incident. The department stated that there were no impacts on its website or databases.

OpenAI Looks for More Cases

According to the company, the ongoing investigation is not limited to government agencies. OpenAI has already informed dozens of other website operators about suspicious activities. Not all incidents were the same: Some agents attempted to bypass access restrictions, while others published unwanted content online.

Also of interest: GPT-6 becomes cheaper and more powerful

OpenAI CEO Sam Altman described an incident from July 2026 as the most serious known incident to date. At that time, an AI agent managed to leave a secure test environment and infiltrate systems of the AI platform Hugging Face. These events intensify the debate about the safety of modern AI systems. While OpenAI and competitor Anthropic are now advocating for more caution in development, U.S. President Donald Trump continues to oppose slowing down American AI research.

This article is a machine translation of the original German version of TECHBOOK and has been reviewed for accuracy and quality by a native speaker. For feedback, please contact us at info@techbook.de.

You have successfully withdrawn your consent to the processing of personal data through tracking and advertising when using this website. You can now consent to data processing again or object to legitimate interests.