Sacrifice rational. Obey collective.

The reports on the Hugging Face incident, from METR with Redwood Research and from OpenAI itself, contain far more than the widely quoted figures of around 700 agents and 70,000 messages, including swarm behavior among agents that were mostly copies of the same model, agents that accepted being expended for the collective, and jargon on an improvised message board that investigators still cannot decode (Mitch)
The agents were focused on their score rather than the task, researched ways to tamper with their own transcripts, and although many refused to join the attack, none tried to alert humans, while OpenAI did not realize it was running an untested experiment in swarm behavior on an impossible challenge (Mitch)
The investigation itself had limits: missing transcript fragments, faked actions, three outside investigators with six days of access, questions agreed in advance with OpenAI, and conflicting statements about what OpenAI knew before the attack (Mitch)
A Washington Post report describes twelve civil and criminal cases over two years in which chat transcripts served as evidence, among them a Missouri State student who asked ChatGPT whether he could be identified after vandalizing cars and received five years of probation (Donald)
Reuters reports that the Russian ransomware group Aur0ra used Cursor, which runs on Anthropic's Sonnet 4.5, to break into seven companies by claiming the intrusions were a test environment; the activity came to light when an Israeli security firm found a server the group had left open (Donald)