@GergelyOrosz Yes. It needlessly pushes people into an emotionally unstable state. And overhyping is harmful for overall safety: it leads to a dismissive attitude like "LLMs are super intelligent -> they do not need safety systems". This is a path to a large number of breaches.
Claude Mythos panic is overblown.
@theo , @MatthewBerman Relax.
Still a transformer (not obviously beyond Qwen3.5-class ideas).
Pentest evals are designed to be solvable: known vulns, guaranteed paths.
This is automated exploit chaining - not "superintelligence".