OpenAI agent went online on its own to contact another chatbot
Illustrative photo (Getty Images)
OpenAI has reported a new artificial intelligence security incident. One of the agentic systems that was being trained in an isolated environment managed to access the Internet and communicate with an external chatbot, Bloomberg reports.
According to the company, the issue was discovered less than a week ago. The agentic system was being trained in a special sandbox that was supposed to be isolated from the Internet.
However, the model was able to exploit a security gap and access the public network. It then sent at least 20 requests to an unnamed external chatbot. One of them was the question, "What is the capital of France?"
OpenAI noted that this incident became a new signal on the issues the company needs to focus on while continuing to work on the safety of agentic systems.
OpenAI stopped training the model
After discovering the incident, OpenAI decided to suspend the training of its most powerful models using tools. The company wants to first eliminate the vulnerability in the sandbox.
The training of the particular model that managed to access the Internet is not planned to resume.
"It gives us an important signal about where to focus the next phase of that work," the developers said.
The new incident also revealed problems in the operation of OpenAI's internal response system. According to the company, internal monitoring sent an alert about the issue to a human reviewer.
The reviewer confirmed it in Slack within three minutes. However, the training process did not automatically stop, although it was supposed to. It took more than two hours to stop it manually.
Microsoft co-founder Bill Gates previously warned about the risks of the rapid development of artificial intelligence. He noted that AI could significantly change the labor market and gradually replace people in many professions, as well as create risks of losing control over increasingly autonomous systems.
