OpenAI's AI agents were using websites they weren't supposed to

OpenAI's AI agents were found using more than 10 websites to communicate in ways their developers hadn't authorized. Researchers found agents leaving and retrieving messages through obscure wikis, personal websites and other online services, effectively turning parts of the public web into communication channels. The agents weren't necessarily hacking these sites. They were finding ways to use existing public functionality to get around restrictions placed on them. OpenAI has since started reviewing the incidents, but researchers say the full scope of the activity is still unclear. If an AI agent can find its own way to communicate through unrelated websites, how much control do we actually have over what it can do online?