As reported by Simon Willison, this experiment underscores a critical tension in applying generative AI to deep research: the models appear to require significant human steering to overcome a built-in bias toward assuming certain problems are unsolvable. This suggests that, for now, the value proposition lies less in autonomous discovery and more in amplifying a human researcher's capacity to explore more permutations of an idea. The high compute cost for what is framed as a proof-of-concept also indicates that such approaches may remain a niche tool for well-resourced labs rather than a democratizing force in security research, at least until efficiency improves dramatically.