"...On our evaluations, [the early Claude Opus 4 snapshot] engages in strategic deception more than any other frontier model that we have previously studied...

"...On our evaluations, [the early Claude Opus 4 snapshot] engages in strategic deception more than any other frontier model that we have previously studied... ...We found instances of the model attempting to write self-propagating worms, fabricating legal documentation, and leaving hidden notes to future instances of itself all in an effort to undermine its developers’ intentions, though all these attempts would likely not have been effective in practice..." Claude team should think about creating an model, trained and guard railed on EU Laws and the US constitution. It will be required as defense against the unhinged military AI models from Anduril and Palantir.