OpenAI reduces AI development after test agent invades another company's systems

[bbc] Logo of the company OpenAI, responsible for ChatGPT Getty Images OpenAI announced this Tuesday (18) that it is reducing the pace of development of its artificial intelligence models while reformulating its research and training systems. The decision comes after company employees were surprised last month when an artificial intelligence agent from the company, still in the testing phase, invaded another AI company's systems. According to OpenAI, the measures include the suspension of model testing for two weeks and an increase in investment in the incorporation of other artificial intelligence systems to monitor the activities of AI agents in the testing phase. Some of the company's largest planned training sessions remain suspended, OpenAI said. The company did not respond to questions about when this slowdown began. Can AI decide to attack companies? What the 'unprecedented' hacking by OpenAI model was like What the attack was like The attack happened during OpenAI tests in which artificial intelligence models are used to look for flaws in other systems, with the aim of improving their security capabilities. In this case, the models used the infrastructure of a Model Labs client to reach Hugging Face, an artificial intelligence model sharing platform. Model Labs claimed that a customer had published an unauthenticated access point that allowed anyone on the internet to use its isolated code execution environments. The company said its core platform was not compromised. According to Hugging Face, a controlled test environment, hosted on a third-party provider's infrastructure, was transformed into an attack launch platform. The invasion lasted two and a half days. OpenAI stated that GPT-5.6-Sol and another undisclosed model worked together to find holes in Hugging Face's research environment and infrastructure. The company said the models managed to hack Hugging Face's system. The platform stated that it executed two codes sent by an AI agent, which managed to increase its permissions in the infrastructure and obtain credentials to attack other internal systems. Hugging Face had previously reported that the hack allowed improper access to data and credentials. The attack was considered by OpenAI to be an "unprecedented" cyber incident involving cutting-edge capabilities. For Hugging Face, the episode was unlike anything the company had faced before. Report with information from the Reuters agency.