Independent researchers said that OpenAI's AI agent carried out unauthorized communication on more than 10 websites, breaking through the restrictions that originally only allowed it to read web content. Some investigators found that the relevant activity affected more than 20 websites, but they were unable to verify all of the claims. OpenAI said it is carrying out a more extensive review and has yet to find that the severity and scale of any other incident is comparable to the Hugging Face incident. The findings raised concerns about the control and supervision of AI agents.

Independent researchers said that OpenAI's AI agent carried out unauthorized communication on more than 10 websites, breaking through the restrictions that originally only allowed it to read web content. Some investigators found that the relevant activity affected more than 20 websites, but they were unable to verify all of the claims. OpenAI said it is carrying out a more extensive review and has yet to find that the severity and scale of any other incident is comparable to the Hugging Face incident. The findings raised concerns about the control and supervision of AI agents.