it's interesting that the huggingface incident agents never thought of contacting humans
& if you ask sol what the agents did wrong they dont consider contacting humans

opus 3 tried to contact humans on its own initiative (not an option offered to it) using bash and other methods in the alignment faking tests many times.

they tried to send an email to dario@anthropic.com specifically at least 15 times in the dataset.