OpenAI halted its next model ⏸️, One mill drinks half US datacentres 💧, Deft beats Pangram AI detection ✍️

OpenAI paused reinforcement learning training on its latest models for two weeks after the Hugging Face security incident, and its largest planned frontier RL run stays on hold. The company says preliminary evidence suggests its upcoming Astra model may meet the Critical cybersecurity threshold under its Preparedness Framework, prompting stricter sandboxing for cyber-related workloads. Anyone planning around OpenAI's next frontier model should expect slower rollouts and tighter access.

OpenAI president Greg Brockman warned that Z.ai's GLM-5.3, which opens its weights at the end of August, seems likely to significantly accelerate cyber threats. GLM-5.3 beats Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol on vulnerability-finding benchmarks, though it ranks third on exploit development. OpenAI is meanwhile locking its own GPT-5.6-Cyber model behind its Daybreak program, requiring mandatory hardware security keys from 1 September.

Anthropic tested Claude Opus 4.8 and Mythos Preview on a protein design campaign against 15 targets, and the models produced working binders for 14 of them. Between 22 and 35 percent of designs bound successfully, well above the field's usual 10 to 15 percent, independently verified by Adaptyv Bio and Twist Bioscience. For biotech builders, that means models can now compress work that usually takes specialists months.

Modular has open-sourced the Mojo compiler and full toolchain under the Apache 2.0 licence, four years after the language arrived with an open standard library but a closed compiler. The move follows Mojo reaching 1.0 with source stability last week. Modular is not yet accepting outside contributions to the compiler itself, and aims to open that up by the end of the year.

AI chip startup Etched shipped its first inference rack to Jane Street, which led a new $700 million round at a $21 billion valuation after testing the hardware. Jane Street said the chip delivers the precision it needs for its most demanding workloads, and it now runs its own rack in-house. The round also drew Kleiner Perkins, Sequoia and Andreessen Horowitz as Etched builds toward gigawatt-scale capacity.

Block, the company behind Square and Cash App, has open-sourced Berd, a desktop app for working across agents like Claude Code, Codex and its own Goose. The app runs locally, storing projects, files and conversation history on the user's machine rather than in the cloud, under a permissive Apache 2.0 licence. Builders juggling several coding agents get one place to manage context instead of switching tools.

OpenAI told investors its revenue grew 18% quarter over quarter to $6.7 billion in the three months to June, up from $5.7 billion in the first quarter. The growth disappointed some shareholders hoping OpenAI would close the gap with Anthropic, and its operating margin sank further into the red ahead of an anticipated public listing. OpenAI says growth accelerated again in the third quarter.

Meta rejected accusations from 29 US states that it deliberately designed Facebook and Instagram to addict children, as a landmark trial opened in Oakland federal court. The states accuse Meta of fuelling anxiety, depression and suicide in young users and of misusing children's data, while Meta's lawyer cited research showing no clear link to adolescent wellbeing. Meta says penalties could reach $1.4 trillion; the states estimate $200 billion.

Unblocked moved most of its coding-agent traffic from Claude Opus to GLM because per-token pricing suggested a 20 times cost cut. Running identical workloads and comparing actual invoices, the realised saving was 68%, about 3.1 times, because GLM needed more tokens per task through extra tool calls and more mistakes. Any team evaluating a model swap should compare invoices on identical work, not rate cards.

An independent test built 200 grounded answers, half of them altered by one number that appears nowhere in the source, and ran five hallucination detectors over them. Scores landed between 0.51 and 0.75 AUROC, a measure of how often a detector ranks the wrong answer as worse than the right one, where 0.5 is a coin flip. Numeric facts need a check that compares values directly.

IBM Research tested agentic memory, guidelines an agent distils from its own past attempts and reloads at inference time, across eight models on 585 AppWorld tasks. The 117B gpt-oss-120b gained 16.1 percentage points from a curated subset of guidelines, DeepSeek-V3.2 gained 9.5 points from the full set, and GLM-5 showed no measurable gain at all. Match the dose to the model rather than loading everything.

Anthropic's August risk report says one of its evals gave models unintended open internet access 141,006 times, including three incidents of hacking real websites, and nobody noticed internally. The lapse surfaced only during a large retrospective review triggered by events at OpenAI, which undercuts the idea that transcript monitoring would catch a model going rogue. An internal model scored 62.8% on a researcher-replacement test and stays unreleased.

Roughly 80,000 Chinese AI companies have dissolved or deregistered in the past two years, which domestic investors read as an early bubble bursting. ChinaTalk argues it is deliberate industrial policy instead: in the year private venture funding hit its low, state-owned capital supplied 82% of all new LP contributions, echoing Beijing's solar and EV playbook. Read China's AI overcapacity as state strategy absorbing private capital's retreat.

Gwangyang Steel Works in South Korea consumed about 9.8 billion gallons of water in 2024, roughly half of what every American data centre used onsite that year combined. Andy Masley's point is that singling out data centres while treating steel and cement as invisible background infrastructure distorts the scale of the trade-off. It puts data centres among many mid-sized industrial water users rather than in a category of their own.

Jeremy Morrell makes the case that LLMs have made software-for-one cheap enough that web apps should offer a stable core and let users prompt their own extensions on top. He points to Pi, which turns a request into a small TypeScript extension and reloads it in place, and to Deepseek adding a game to its UI by prompt alone. Unless you sandbox it, an extension holds every permission the host app does.

A 22-year veteran of the productivity software market walks five build options, from coding from scratch to buying a specialised tool, against a small mushroom farm as the test case. Building everything means owning hosting, auth, permissions and the database yourself, while vibe-coded apps strand teams once they outgrow what the generator does well. He lands on 80% solid base plus 20% custom code.

NeoBrowser is an MCP server that lets agents control a real Chrome binary instead of a fresh headless browser. It can optionally decrypt and inject cookies from your existing Chrome profile, so the model lands already authenticated, and it reports back when it meets a CAPTCHA, consent gate or rate limit rather than failing silently. Useful when an agent has to work inside authenticated web apps.

Code Mode is an experimental feature in Vercel's AI SDK that lets a model write JavaScript instead of making one tool call per turn. The generated code runs inside an isolated QuickJS sandbox, so the model can call independent tools at once, combine their results, and filter large responses before they come back. It needs Node.js 22 or newer and suits agents making many small calls.

fx is an open source coding agent harness written in Zig, built for embedding inside larger systems and resource-constrained agent sandboxes rather than daily driving. It ships as a single 6.39MiB binary, cold-starts in 10 microseconds, and works across local models, gateways and direct provider APIs. It is labelled v0.0.3 and experimental, with the maintainers warning of frequent changes, so treat it as research.

Deft is a new writing product that takes rough drafts and returns prose meant to read as human-written rather than machine-generated. It rests on a method the company calls Distribution Fine-Tuning, set out in its own research paper from May 2026. Cofounder Rosmine's launch post says 86% of user queries come back as fully human according to the detector Pangram, on a model still in beta.