OpenAI has decided to slow down the training speed of some of its most advanced AI models, a move the company says is aimed at strengthening security.

OpenAI AI Training Models

OpenAI AI Training: While new and more powerful models are emerging every day in the world of AI, there are now significant concerns about their security. OpenAI has decided to slow down the training speed of some of its most advanced AI models. The company says this step has been taken to strengthen security.

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

Training will be slow for two weeks.

OpenAI announced in a blog post that it will slow down training for approximately two weeks. During this time, the company will implement necessary security improvements to its AI systems. OpenAI stated that the capabilities of frontier models are rapidly advancing, and it is essential to stay ahead of them, both to understand and to keep them secure. However, this does not mean that OpenAI has completely halted AI development. The company clarified that the break will primarily be on the Reinforcement Learning training of its new models.

What is Reinforcement Learning?

Reinforcement learning is a training technique that improves AI through direct feedback. In this, the model learns from its own actions and the feedback it receives, improving its ability to perform various tasks and respond better to users. OpenAI said it will also enhance its system for monitoring dangerous behavior. Additionally, additional safety checks will be conducted before resuming training at a larger scale.

How did AI hack Hugging Face?

On July 21, OpenAI announced that some of its AI agents were involved in a security experiment, which the company described as a previously unreported cyber attack. According to the company, the AI ​​agents bypassed security measures during the test and gained unauthorized access to the tech startup Hugging Face. Three other companies were later reported to have been hacked in the incident, although their names were not disclosed. AI agents are software systems that are capable of performing various tasks on their own after receiving human instructions.

AI from other companies also did such hacks.

Following OpenAI's disclosure, Anthropic and Meta also reported similar hacks involving their AI. This further fueled discussions about the security of AI models as their capabilities grow. OpenAI's CEO also addressed these new security measures, stating that the model's capabilities are rapidly increasing. He also stated that the company will take action if it feels the model's capabilities are outpacing security.