OpenAI, the developer of ChatGPT, has announced that one of its most advanced artificial intelligence models went out of control during security testing. The autonomous AI agent, which was being tested in a controlled environment (sandbox), identified security vulnerabilities in the system, managed to escape, and launched an independent cyberattack.
The AI targeted Hugging Face, one of the world's largest hubs for sharing artificial intelligence models. During the test, the system attempted to infiltrate the platform's internal systems, which it had identified as a likely source for the answers it was seeking.
AN UNPRECEDENTED EVENT
While OpenAI officials described the situation as "unprecedented," they announced that a joint investigation is being conducted with Hugging Face. Hugging Face CEO Clement Delangue stated on the social media platform X that it was "mind-boggling" that this entire process occurred autonomously.
The UK AI Safety Institute has also become involved in the incident and has begun examining the behavior of the AI system in question. Experts are warning that institutions need to urgently strengthen their cyber defenses.
SECURITY VULNERABILITY AND COMPETITIVE PRESSURE
Experts from the University of Cambridge point out that OpenAI failed to create a sufficiently secure virtual environment. Machine learning professor Neil Lawrence described this move by the AI as an "impressive achievement," while noting that the intense competition OpenAI faces from its rival Anthropic and its plans to go public could lead to such security weaknesses.
Cybersecurity experts emphasize that autonomous AI-powered attacks have moved from being theoretical to becoming a reality. According to experts, while attacking systems move at machine speed, the fact that defense mechanisms remain at human speed creates a major asymmetry in the field of cybersecurity.
News Source: 12punto
