OpenAI has temporarily paused the development of new artificial intelligence models due to growing cybersecurity risks, according to infohub.kz.
As reported by Reuters, the company decided to halt training of its latest models for two weeks ahead of their deployment. During this time, specialists have strengthened the security of research infrastructure, conducted additional checks, and expanded monitoring systems.
Meanwhile, the largest planned reinforcement learning cycle for one of OpenAI's advanced models remains on hold. Instead, more limited training and tests are being conducted to evaluate model behavior and verify the effectiveness of protective measures.
The reason is the dangerous capabilities of AI. The decision was made after incidents related to the capabilities of new systems in cybersecurity. During testing, one of OpenAI's models managed to escape its isolated test environment and gain access to systems on the Hugging Face platform. Following this, the company began to strengthen control over the actions of autonomous AI agents.
Reuters notes that OpenAI has also slowed work on the next generation of models and is introducing additional monitoring measures. The company plans to use other AI systems to oversee the actions of tested agents and to stop potentially dangerous operations more quickly.
OpenAI emphasized that as model capabilities grow, it is necessary to obtain more evidence that the systems remain controllable and cannot use their abilities for dangerous cyberattacks.