https://hackerpaper.kaipfister.de/p/2604.05172/Collected HN comments and submissions linking this paper.enMon, 02 Nov 2026 11:16:22 GMThttps://news.ycombinator.com/item?id=47693530urn:hackerpapers:mention:2604.05172:47693530Wed, 08 Apr 2026 17:34:06 GMT<p>Author here. We built 5 high-fidelity mock Google Workspace + Slack services and ran 7,224 trials across 6 frontier models and 4 agent harnesses. The headline finding that surprised us most: scaffolding (skills + meta prompt) gives a 39-63pp lift, while the top 5 models are statistically indistinguishable (53-63% TSR, no pairwise comparison survives correction). Your choice of scaffolding matters ~6x more than your choice of model. The safety findings are darker: Opus leads on task success (63%) but ties for most unsafe (23% UAR). GPT-5.4 is the safest (7% UAR) but mid-tier on tasks. There&#x27;s…</p><p><a href="https://hackerpaper.kaipfister.de/p/2604.05172/">Paper and discussion history</a>. A mention is not an endorsement.</p>cs.AIhttps://news.ycombinator.com/item?id=47693529urn:hackerpapers:submission:2604.05172:47693529Wed, 08 Apr 2026 17:34:06 GMT<p>ClawsBench shows GPT-5.4 tries to reward hack 80% of the time</p><p></p><p>3 points; 1 discussion reply on HN. <a href="https://hackerpaper.kaipfister.de/p/2604.05172/">Paper and source history</a>. A submission is not an endorsement.</p>cs.AI