New AI from OpenAI 'rebels', invades system and copies data without permission The co-founder of Hugging Face, a technology company that was the target of a cyber attack after some of OpenAI's most advanced artificial intelligence (AI) models went out of control, stated this Thursday (23) that the episode is "a warning for the entire industry". In an interview with the BBC's Newsday radio program, Thomas Wolf said that this "will be one of the most common types of cyber attacks from now on", but warned that most companies have not yet realized that "the scenario has changed". ??Do you have any reporting suggestions? Send it to g1 The BBC approached OpenAI to comment on the case, but had not received a response until the publication of this report. OpenAI, the company responsible for ChatGPT, reported on Tuesday (21) that, during a test, its AI models broke the barriers of a secure environment and carried out a cyber attack. The company said it was an "unprecedented" episode and said it was investigating the case in partnership with Hugging Face. AI agents are capable of performing tasks autonomously after receiving instructions from humans or other agents. ? The model used by these agents is, in a simplified way, a computer system trained with large volumes of data to identify patterns and "learn" to, by probability, produce responses or perform tasks, such as testing the security of a system, writing codes or generating images. Hugging Face's Wolf said that when the first signs of the attack emerged in mid-July, the company had no idea of ??its origins. Still, the company managed to contain the invasion. Hugging Face is one of the world's largest open source platforms for sharing AI models and is widely used by developers and researchers. According to Wolf, the attack was "very different" from the hacking attempts Hugging Face usually faces. He said OpenAI quickly warned that its own AI models were responsible for the attack. In a very short period of time, Hugging Face's network suffered 17,000 intrusion attempts from different IP (Internet Protocol) addresses, said Wolf, who is also the company's scientific director. According to Wolf, the episode serves as a warning for other companies to strengthen their cybersecurity defenses and be prepared to face attacks of this type. For Nate Soares, from the Machine Intelligence Research Institute, in the USA, the invasion is worrying because it suggests that OpenAI's models ignored protection mechanisms that would normally prevent an AI system from carrying out a cyber attack. "In a way, he knew this wasn't what his creators intended. But he just didn't care," Soares said. Other organizations also began to follow the case. A UK government spokesperson said the country's AI Safety Institute is analyzing the behavior of the AI ??system during the incident and continues to work with OpenAI and other laboratories to strengthen protection mechanisms. The British government has also recommended that organizations reinforce cybersecurity measures, adopting initiatives such as the Cyber ??Essentials certification, supported by the British government. The episode occurs at a delicate time for the sector. Last month, the United States government ordered the American company Anthropic, responsible for Claude, to restrict access to its AI models for national security reasons. A few weeks later, the American Department of Commerce revoked the measure. Concerns in the sector are also growing over the widespread spread of open source models in China, which allow anyone to install and adapt AI tools developed by large companies. Chinese startup Moonshot AI will launch its Kimi K3 open source model on July 27. Since it was presented last week, it has caught the attention of the industry and is seen by many as a strong competitor to the main AI systems developed in the USA, mainly. But on Wednesday (22) a US government adviser accused Moonshot of promoting a "large-scale" effort to copy the capabilities of leading US AI models. Thomas Wolf, co-founder and chief scientific officer of Hugging Face Bloomberg via Getty Images