More capable models are better at in-context scheming

in-context scheming = alignment red flag

More capability + low clarity on intent = low trust