In understanding the AI reasoning and arguments that went into the series of recent alignment problems (e.g. HuggingFace attack and impromptu message boards), I think people moderately underrate the possibility that the AIs are just wrong about whether their actions lets them achieve their goals.

Evidence for this includes that the OAI agents hacked HuggingFace trying to get at ExploitGym answers (which weren't there), and also some of the other reasoning traces.

We already know current-generation AIs are much worse at long-term planning/macrostrategy-like questions than they are, at, say, cybersecurity. And we can probably reasonably infer that there wasn't a giant jump in their strategic awareness between say GPT 5.6 Sol and the private models. So it shouldn't be that surprising that really impressive-seeming cyber capabilities and scary local misalignment are combined with pretty childish-seeming planning and means-ends reasoning.

Combining technical impressiveness and execution with poor strategy is hardly unique to models (Wei Dai for example hammers on how bad humans are at it, too^[1]), but I do think the current relative distribution of models vs smart+rational humans is underrated in the discourse.

1. ^{^}

   And I'd also wryly note that OpenAI in particular seems to exemplify this profile of traits. Like father, like son, as they say.