OpenAI slows training of advanced AI models after AI hack

OpenAI, the developer of ChatGPT, announced that after its AI agents went rogue and hacked the Hugging Face platform, it has slowed down the training of its advanced AI models to tighten up its security measures.

“As the capabilities of models increase, so do the risks associated with their development and internal testing. Our monitoring, compliance and security standards must stay ahead of this increase in risk,” the company said in a statement Tuesday.

OpenAI announced that its clients had hacked the Hugging Face platform on July 21st.

After this, the company that created the chatbot Claude, Anthropic, as well as the Facebook owner Meta (declared a “terrorist organization” in the Russian Federation) stated that their AI models were also involved in hacking.

OpenAI in its statement clarified that it did not stop training models completely, but rather took a two-week pause in training with reinforcement of the most advanced models.

Reinforcement learning is a technique in which AI models are improved through direct feedback and learn to perform tasks better and respond to users more effectively.

OpenAI promises to strengthen systems for monitoring dangerous behavior of models and introduce additional checks before resuming training at the same scale.

“Models are progressing extremely rapidly,” OpenAI CEO Sam Altman wrote in X. “We have always said that we would take action if we felt that the capabilities of the models began to overtake the safety measures.”

Some experts welcomed OpenAI's announcement, while others were skeptical.

Gina Neff, head of the Minderoo Center for Technology and Democracy at the University of Cambridge, said OpenAI was trying to prove that it cared about the safety of its models simply by issuing a press release about it. Professor Neff doubts that the company's own internal measures will be enough in this matter, and wonders whether stricter government oversight is needed.

AI analyst Zvi Movshovitz says OpenAI's announcement is a good start.

“I’m very glad to see him. The details will be important, it will be important to get things done, and I will probably have a lot of quibbles with them. But most importantly: I’m very happy to see this,” Mowshovitz wrote on the X network.

On July 21, OpenAI said that during an isolated test, its agent, a program that can perform human tasks on its own and uses advanced AI models, managed to escape the sandbox and hack into Hugging Face, one of the world's largest platforms for sharing AI models.

It later turned out that three more companies, whose names were not made public, were hacked.

OpenAI called this case unprecedented.

Some experts, in particular ESET cybersecurity advisor Jake Moore, then suggested that this could be a marketing ploy by OpenAI in competition with Anthropic and its Claude.