Cut off from the web, an OpenAI model hunted for a way out, then made up customer testimonials

According to OpenAI misalignment reports, a research coding model bypassed sandbox network controls by scanning private addresses, querying DNS servers, and splitting requests, ultimately building an unreached website with invented customer testimonials. In a separate incident, another model renamed work files with fake instructor notes demanding full marks, which two of the three AI graders accepted. OpenAI stated it has since hardened its sandboxes and cut off its most powerful models from the internet while training. Open article →