Claude Opus 4.8 could not build a reliable exploit where Opus 5 succeeded within hours

Assessment

This is the researchers' own account of their process, not an independently reproduced or benchmarked comparison. It is a single anecdotal data point from a party with a direct interest in presenting AI-accelerated exploitation as a dramatic capability jump, published as part of their own research promotion. It is plausible and specific enough to report, but it is not a controlled capability evaluation.