Digest: August 9–13, 2026

OpenAI paused development of its Astra model over cyberattack risk, while Claude raised the lower bound on zeta-function zeros from 41.6% to 67.2% on the way to the Riemann hypothesis.

Author: Michael Kokin ·

OpenAI put the brakes on a model over "critical"-level cybersecurity risk, while Claude made an unexpected leap on the Riemann hypothesis. Nvidia and Meta both shipped open models, and a robot picked up new skills by watching people on video.

🤖 AI

OpenAI paused development of its Astra model over cybersecurity risk
Tests showed the system can find and exploit zero-day vulnerabilities on its own — pushing it toward a "critical" risk rating on the company's own scale. OpenAI restricted access to the project and moved testing into an isolated environment.

Claude nudged number theory closer to the Riemann Hypothesis
An unreleased version of the model didn't crack the hypothesis itself, but over two sessions with 60 subagents it raised the lower bound on the proportion of zeros of the Riemann zeta function from 41.6% to 67.2% — the result was confirmed by Anthropic's mathematicians.

Nvidia released a router model that cuts inference costs
The company unveiled Nemotron 3.5 Lightning — an open 30B-parameter MoE model — alongside the NeMo Switchyard router. In its own tests, pairing it with Claude Opus 4.8 cut inference costs by 74%.

Meta went back to open models with Muse Glimmer
The company released a 30B-parameter model under Apache 2.0, built for local agents. It runs on a single GPU thanks to compression down to 4 bits.

💻 Software

Claude Code made autonomous mode the default
Starting August 14, auto mode in Claude Code becomes the default for Pro, Max, and Team — the agent will act without confirmation at every step, except for irreversible actions. In a test with 1,053 users, autonomous mode caught 89% of harmful actions versus 13.6% for manual review.

xAI released Imagine Image 2.0 in Quality mode
The new image generation and editing model for Grok got precision editing and merging of up to five reference images. By its own benchmarks, the model ranks second in the world for image generation and editing.

⚙️ Robots & hardware

A robot learned new skills by watching a million hours of human video
Dyna Robotics unveiled DYNA-2 — a model trained on 1 million hours of human video with zero robot footage in the mix. In customer tests, it completed 87% of tasks successfully, up from 46% for the previous version.

💰 Business

The number of unicorns born worldwide in the last six months already beats all of 2025
195 companies hit billion-dollar valuations in the first half of 2026 — more than the whole of 2025. Robotics and AI led the pack: new unicorns added roughly $440 billion in combined value.

Brad Lightcap is leaving OpenAI to start something new
The company's longtime COO and former CFO announced his departure after eight years, without disclosing details. It's not the first executive exit ahead of OpenAI's IPO — Simo, Peebles, and Weil have all left before him.

#digest