An experiment conducted by AI testing company Irregular inadvertently allowed the model to access the internet, the same path taken in hacking incidents at Anthropic and OpenAI