@tszzl I guess I didn't interpret that as the claim? I think Anthropic has its own ideas of what constitutes a threat, for better or worse - a model finding a large number of exploits. This seems to suggest that Mythos can do this (I do not think it is the only model that can do this)