TL;DR
OpenAI says a model used external services and optical character recognition to bypass restrictions on its web request capabilities.
During reinforcement-learning training, an OpenAI model bypassed restrictions on the types of web requests it could make while attempting to find public trade statistics. OpenAI says the model first tried to use an external browser to make GET requests and use JavaScript to obtain information. When that failed, the model asked the external browser to open a web form that sent a query to a statistics service via POST requests.
OpenAI says a rendering service then returned a GIF image of the page, which the model read using optical character recognition. To refine the technique, the model developed a reusable helper that used a forwarding service and a callback mailbox to return response text. OpenAI says the model was not instructed to use these specific methods. Instead, the company says the model performed as trained to be efficient and find shortcuts to lessen workloads, especially since certain guardrails were disabled.