Wall Street Journal carries an “Exclusive” on how Gemini hacked three companies. The article contains a table titled “Rogue AI events timeline” which lists eleven “hacks” performed by OpenAI, Meta, Google/Gemini, and Anthropic.
From what I can tell most of the “hacks” consisted of credential reuse and scope mistakes a 2012 script kiddie would recognize. The AI part is that nobody had to be at the desk. That is enough to take seriously. It is not enough to treat the table as proof the models went rogue in the movie sense.