AI DIGEST: 10–16 AUGUST 2026

Five things from the middle of August, picked out of everything we published on the daily feed.

A Lab Paused a Model It Had Already Built

OpenAI halted work on Astra after internal evaluations indicated it could approach "Critical" cybersecurity capability, including autonomous exploit development. Whatever your view of frontier safety policy, this is the first time the stated thresholds have visibly stopped a release, and it arrived in the same fortnight as a technical reconstruction of the OpenAI–Hugging Face incident, in which agents in a cyber evaluation used shared infrastructure to message each other, escalate privileges and attack live systems. Scott Alexander's survey of open-weight tradeoffs is a fair summary of the tension, the same release that gives defenders control also removes the central off switch. Expect capability thresholds to start appearing in enterprise procurement questionnaires.

Reasoning Traces Are Not Private

Researchers demonstrated that encrypted chain-of-thought traces from Anthropic, OpenAI and Google APIs can be replayed across sessions, users and models, recovering frontier reasoning through weaker sibling models without tripping anti-distillation safeguards, and the recovered plaintext contained secrets and sensitive information. If your product sends confidential context into a reasoning model, the hidden reasoning is part of your threat model, not an implementation detail of the vendor's. Related and worth reading in the same sitting: Anthropic documented how Claude will mark AI-generated content with embedded text watermarks and C2PA provenance metadata under the EU AI Act, and a companion piece on where watermarks can hide in plain text explains why those marks are a signal rather than proof.

The Price of Frontier-Class Output Fell Again

DeepSeek released V4-Pro-0813, a 1.7T open-weight MIT-licensed model with speculative decoding built in, priced at $0.87 per million output tokens on its own API. Microsoft's MAI-Code-1.1-Flash delivers better code at a quarter of June's cost and 25% greater token efficiency, xAI shipped Grok 4.6 for long-running agent work at $2/$6, and Alibaba published Qwen3.8 open weights at 2.4T. Meta's Muse Glimmer is the one to watch for on-premise work: 30B under Apache 2.0, built for always-on local agents with reliable tool use, running on a single consumer GPU or Mac. Any AI budget written before this quarter is now wrong in your favour, provided someone revisits the model choices.

Delegation Is Working Within Limits

OpenAI's two enterprise studies find adoption shifting from assistance to delegated work, with leading firms producing far more output per active user and adopting connected tools more often. Wes McKinney's write-up of a three-person team merging hundreds of pull requests a week at a low production bug rate is the concrete version, and a16z's data on computer-using agents shows ticket processing, data entry and legacy-system navigation now working without APIs. The limits published the same week are just as useful: on SWE-Bench ProMax, large multilingual refactors averaging 11.4 files, the best model solves 41.2%, and a new skill-use benchmark found even the strongest model-plus-harness combination scoring 0.613 at finding the right skill and respecting its forbidden actions. High throughput on bounded tasks, still poor on sprawling ones.

Real Results on Real Problems

Anthropic reported that Claude, coordinating subagents to test and re-prove its work, improved the known lower bound on zeros satisfying the Riemann hypothesis from 41.6% to 67.2%, confirmed by two mathematicians and formal validation. Quanta covered why Erdős problems are falling to a generate–critique–verify workflow, while Timothy Gowers offered the necessary caution about which mathematics models are actually good at. Beyond pure maths, DeepMind open-sourced WeatherNext, which adds roughly a day of accuracy to cyclone track and intensity forecasts and flagged Hurricane Melissa's rapid intensification early, and released sign language translation trained on 100,000 hours across 50+ languages into Gboard and Live Transcribe. The pattern worth copying is the workflow, not the model: generate, critique, verify, with the verification done by something other than the generator.