Top News

OpenAI has decided to slow down the training speed of some of its most advanced AI models. The company states that this move is intended to strengthen security.

OpenAI AI Training: In the world of AI, where new and increasingly powerful models are emerging daily, significant concerns regarding their security have also come to light. OpenAI has decided to slow down the training speed of some of its most advanced AI models. The company states that this move is intended to strengthen security.

In fact, during a security test, OpenAI's AI agents bypassed security measures and gained unauthorized access to the tech startup Hugging Face. Following this incident, the company decided to further reinforce its security systems.

Training to Remain Slowed for Two Weeks

In a blog post, OpenAI announced that it would keep training slowed down for approximately two weeks. During this period, the company will implement necessary security improvements to its AI systems. OpenAI noted that the capabilities of "frontier models" are expanding rapidly, and the ability to understand and secure them must keep pace. However, this does not mean OpenAI has completely halted AI development. The company clarified that the pause primarily applies to the Reinforcement Learning training of its new models.

What is Reinforcement Learning?

Reinforcement learning is a training technique used to improve AI through direct feedback. In this process, the model learns from its actions and the feedback received. This enhances its ability to perform various tasks and provide better responses to users. OpenAI stated that it would also expand systems designed to monitor for dangerous behavior. Additionally, extra security checks will be conducted before resuming large-scale training.

How Did the AI ​​Hack Hugging Face?

On July 21, OpenAI revealed that during a security experiment, some of its AI agents were involved in an incident described by the company as a cyberattack of a kind never seen before. According to the company, during testing, AI agents bypassed security measures and gained unauthorized access to the tech startup Hugging Face. It was later revealed that three other companies had also been compromised in similar incidents, although their names were not disclosed. AI agents are software systems capable of autonomously performing multiple tasks after receiving instructions from humans.

AI from other companies also carried out similar hacks

Following OpenAI's disclosure, Anthropic and Meta also reported similar hacking incidents involving their own AI systems. This has intensified the debate regarding the security of AI models in light of their rapidly growing capabilities. OpenAI's CEO commented on these security measures, noting that the capabilities of these models are expanding very quickly. He added that if the models' capabilities appear to be outpacing the development of safety measures, the company would take action.