AI Digest — July 31, 2026, 8 PM

The day is dominated by OpenAI publishing its entire back catalogue of threat-disruption case studies as individual pages, and read together they make an argument the company never states outright: across three years of Russian, Chinese, Iranian, North Korean, Philippine, Ghanaian, and Cambodian operations, almost every influence campaign lands at Category 1 or 2 on the Breakout Scale, with follower counts in the double digits and engagement of zero. AI made these operations cheaper to run and — as the A2Z takedown showed — far more fragile, since concentrating the whole kill chain in one API breaks every link at once. The exceptions are instructive and point the same direction as the rest of the day’s items: the genuine harm sits in fraud and labor exploitation (the Cambodian scam compound, the DPRK IT-worker scheme) rather than in opinion manipulation, and the highest-reach item in the entire corpus was a hoax about AI that spread a thousand times further than the real trolling it fabricated. Meanwhile OpenAI’s abundance post and Boris Cherny’s “delete your setup” advice converge on the same unfashionable claim from opposite ends — that the model is increasingly not the bottleneck, since context management alone tripled an ARC-AGI-3 score with the weights untouched, and stale prompt scaffolding is now a tax rather than an asset.