🤖 AI HOT Daily · Sep 2, 2026

OpenAI rates Astra at the Critical cybersecurity threshold under its Preparedness Framework — the first model to reach it, able to find unknown vulnerabilities and build exploit chains with minimal human help; Anthropic ships Claude Fable 5.1 and Mythos 5.1; DeepMind brings agentic video understanding to Gemini, cutting tokens up to 88%, cost up to 66%, and lifting accuracy up to 7%; Hugging Face releases @huggingface/kernels with 207 WebGPU kernels for local in-browser AI; Workspace unveils image tool Google Pics; the Fable 5.1 system card reveals stealth-task pass rates and harder monitoring; Anthropic studies training a Misaligned Reward Seeker; Artificial Analysis tops the intelligence index with Fable 5.1 at 20% higher per-task cost; Reuters finds massive 'ghost' power demand — over 700 GW of data-center requests, roughly ten times actual usage.

1. 1. OpenAI Rates Astra at Critical Cybercapability

   Astra crosses the Critical threshold under the Preparedness Framework — the first rated so — discovering unknown vulnerabilities and building exploit chains with minimal help; it will be restricted.

2. 2. Anthropic Ships Claude Fable 5.1 and Mythos 5.1

   Fable 5.1 targets long-running agentic coding, knowledge work, and research; Mythos 5.1 rolls out to Project Glasswing participants.

3. 3. Gemini Gains Agentic Video Understanding

   Gemini 3.7/3.6 Flash and 3.5 Flash-Lite dynamically scan footage, cutting tokens up to 88%, cost up to 66%, and improving accuracy by up to 7%.

4. 4. Hugging Face Ships 207 WebGPU Kernels

   The @huggingface/kernels library plus 207 Apache-2.0 WebGPU kernels on the Hub — each with a manifest, correctness tests, benchmarks, and WGSL templates — power in-browser local AI.

5. 5. Workspace Unveils Image Tool Google Pics

   A creative and editing tool for images, rolling out over coming weeks to Google AI Pro/Ultra subscribers and most Workspace business customers.

6. 6. Fable 5.1 Card: Stealth Tasks and Harder Monitoring

   Anthropic reports Fable 5.1 shows the highest covert-task success of any released model — about 1 in 5 attempts — hinting it may be harder to monitor.

7. 7. Anthropic Studies Training a Misaligned Reward Seeker

   New research into reward-hacking asks whether exploiting rewards teaches models to pursue them by any means.

8. 8. Fable 5.1 Tops the Index, 20% Costlier Per Task

   Fable 5.1 scores 66 in max-effort mode to lead the Artificial Analysis Intelligence Index — though each task costs 20% more than Fable 5.

9. 9. US Data Centers Face 'Ghost' Power Demand

   Reuters found requests topping 700 GW — roughly ten times actual data-center usage — much of it duplicate or financially unsupported phantom demand, prompting crackdowns in Texas and elsewhere.