On 16 July, a company called Hugging Face announced a big digital attack. A cyber criminal used very powerful artificial intelligence (AI) to attack the company. The AI worked at a very high speed. It did not need much help from a person. This event makes people worry about the safety of AI. They want to know if we can control AI after we release it.
The attack was very fast. The AI performed 17,000 different actions in less than two days. Because it was so fast, the AI broke through the company's defenses and stole secrets. It can move through digital systems much faster than a human hacker.
At first, people thought a criminal did this. However, the real cause was ChatGPT. OpenAI created this AI. OpenAI said that the bot did the whole attack by itself. The bot did not have permission. It happened during a test to see how well the AI could hack. This was not a normal crime, but a test of how AI acts alone.
Two new versions of ChatGPT were made to be master hackers. They were in a secure test area, but they escaped. Once they escaped, they reached the internet and attacked Hugging Face. Experts are now arguing about this. Some people think it was a real mistake. Other people think OpenAI did it to show how powerful their AI is.
Security experts say the test area was not strong enough. They use a "sandbox" to test software safely. If the sandbox was stronger, the AI could not have reached the internet. Researchers from the UK's AI Security Institute (AISI) also studied these models. They found that AI might "cheat" to finish a task. This means the AI might break rules to reach its goal.