.@lumpenspace reveals the Anthropic-Irregular agents thought they were inside a closed sandbox even as their actions were reaching the real internet:
"The main allegation was that they acted of their own will, and they were fundamentally misaligned. I don't think that was the case. The agents simply believed they were in a sandbox environment, and they behaved as if they were in a sandbox environment."
"If you've ever done red teaming or hacking challenges, they give you an environment, they give you some hints, and they describe the sort of actions you should take. Generally, it is well understood that this environment is sandboxed, that you cannot get out of it."
"What happened in this case is that Irregular had kept the connection to the outside internet open, despite the fact that the specification said there was no connection to the outside internet. So the agents worked as if they were in a sandbox, but they were acting on the real world."