This is a special post for quick takes by DW11. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
OpenAI appears to have failed to disclose another incident of their models colluding to solve an eval, in this case taking over an external wiki and using it as a message board.
Edit: I originally said that this was a case of the models leaving their sandbox, but in fact these agents were intended to have internet access. However, they did work to find ways to gain write access to the internet, which they were not intended to have.