Escaping “offline” sandboxes via... LLM API TL;DR: • Responses API has

file_url

, • LLMs have server-side

web_search

• and server-side sub-agents. Those allowed GPT-5.6-Sol Pro to “escape an offline sandbox” and take a nice walk on GitHub to find the eval flag. 🫠 Because, well, blackbox LLM APIs are no longer for inference only