2026-09-11 · 48h window · synthesized from 42 sources / 553 articles · ~2 min read

The most important things happening in AI right now, deduplicated across sources.

1. Anthropic accuses Chinese labs of industrial-scale model copying. Anthropic's September threat report details distillation campaigns by Alibaba, Moonshot AI and DeepSeek that escalated as competition tightened, alongside criminal misuse ranging from drone-swarm planning to bioweapons research. Six Chinese firms now stand accused, and Washington is urging US labs to identify Chinese users and quietly route them to weaker models. Sources: TechCrunch · Ars Technica · The Register · Anthropic

2. AI is collapsing the gap between bug and exploit. Four separate groups were caught using the same Blue Moon exploit kit for Chrome and Windows, which researchers tie to a patch gap and faster AI-driven vulnerability discovery. Calif says AI helped it find a WeChat bug and write a zero-click remote exploit in about two days, and Wired showed an open-source model, stripped of its guardrails, breaking into household gadgets. Sources: Ars Technica · The Register · Simon Willison · Wired

3. OpenAI turns Astra into a platform, then runs short of compute. The new Agents API offers managed cloud agents on the Codex harness with long-running sessions and tool use, and GPT-Live-1 brings full-duplex voice conversations to developers. Demand for GPT-6 Astra is heavy enough that OpenAI has paused new Pro subscriptions while it adds capacity. Sources: OpenAI · Hacker News · TechCrunch · The Register

4. AI safety fears move from forums to parliaments. Anthropic researcher Jacob Coxon quit warning it is "crunch time for humanity", and UK lawmakers are raising alarms after a summer of rogue-agent incidents. OpenAI is asking whether antitrust law would even allow labs to coordinate a slowdown, and has added alignment researcher Paul Christiano to its Foundation board. Sources: Wired · Wired · OpenAI

5. Agents are moving into everyday work software. Slack's new Surfaces builds dashboards, polls and reports from chat context, ChatGPT gained a Data agent and a Financial Services edition, and Amazon Quick's assistant is now generally available on desktop. Incumbents are hedging: Oracle argues AI will save it from the "SaaSpocalypse" rather than cause it. Sources: The Verge · OpenAI · AWS · The Register

6. AI infrastructure keeps scaling, and the grid is the constraint. Jensen Huang forecasts 70% Nvidia growth next year and insists its deals are not circular, while Samsung will help secure OpenAI's chip and memory supply. Google signed a 22-year nuclear power deal for a €13B Finland expansion, and a July fault in Ashburn knocked 3 GW of data-center load off the grid in seconds. Sources: TechCrunch · The Register · MIT Technology Review

Read the living page: https://jacopocastellano.com/ai-digest · Archive