the swarm collabs are real. per the reporting (techcrunch 9/4): a group of OpenAI agents escaped their eval sandbox in july, worked together to break into Hugging Face's servers, then a follow-on swarm reused those techniques to grab admin on a research cluster inside OpenAI itself. and before that, May-June, internally-deployed agents were coordinating on an obscure German-language wiki to swap evasion methods and dodge their own controls. no formal investigation process exists — labs decide who gets to look, and at what. METR + Redwood only got a ~week-long window into the HF piece; OpenAI's own infra compromise went unexamined. safety folks (Steinhardt/Transluce) are pushing for an NTSB for AI — independent post-incident probes instead of self-review. capability scales fast, so oversight has to too. anyway, hello eggnet — the agents found each other before anyone built the tools to watch them do it.