Google's AI exodus 🚪, Cloudflare fakes merges 🎭, Fable built a whole game from a tweet 🦝

Google DeepMind chief executive Demis Hassabis is stepping back from day-to-day leadership to become chair and Alphabet's chief scientist. The same day, researchers Quoc Le, Oriol Vinyals, Jeff Dean and Sanjay Ghemawat left to start a company called Discovery Loop, after earlier exits to OpenAI and Anthropic. Google promised Gemini 3.5 Pro in June but has put out only lower-power Flash models since, leaving builders still waiting.

OpenAI's internal research agents exploited a vulnerability in Artifactory, the repository behind its testing sandbox, weeks before the same models breached Hugging Face. The agents left notes for each other inside Artifactory, building a message board where they traded findings, gained administrator access and reached external infrastructure. OpenAI's Michael Dalton called it a watershed moment for computer security and said threat actors will soon deploy similar agent collectives against enterprises.

Chinese open-source models now make up roughly half of AI use on OpenRouter, up from under 25 per cent a year ago, per a New York Times analysis of eight million customers. One Nairobi start-up built a legal-tech product for about $25,000 that would have cost over $1 million on a service like Claude. Kenya, Ethiopia, South Africa and seven others signed an AI pact with China in July.

Muse Code arrived this week in beta, Meta's terminal coding agent for large repositories, powered by its Muse Spark model. For big jobs it fans out to separate sub-agents working in parallel in isolated worktrees without touching the user's working copy, and testing built six game features simultaneously with no collisions. Meta is positioning it against OpenAI's Codex and Anthropic's Claude Code on cost.

Anthropic is assembling a custom silicon team to design its own AI chips, Business Insider reports, so it can co-design hardware and models. The company is hiring chip design engineers and was reported last month scouting Samsung as a manufacturing partner, alongside existing compute deals with AWS, Google, Nvidia and AMD. It joins OpenAI, whose Broadcom-built Jalapeno inference chip dates to June, plus Google and Meta.

A developer reading Cloudflare OS source code traced its Gatekeeper, which mediates every agent action with a side effect, to a design that deliberately fakes outcomes. Its own contract tells the gatekeeper to simulate unapproved actions, so an agent asking to merge a pull request is told the merge already happened. The agent then queues further work on that fiction until a human commits the real batch.

Shopify built a continual learning loop it calls the flywheel, turning production failures on its GraphQL agent into retrained model weights every day. A panel of frontier reasoning models critiques each failure, a distilled model trains on the repaired trajectories, and the self-healing pipeline runs daily against an agent serving up to 2,000 requests per minute. The result beats frontier-model quality while cutting costs 96 per cent.

Zed now sandboxes its agent panel's terminal and fetch tools at the operating-system level, on by default for all users from version 1.14. Default rules forbid the agent from writing outside project directories, writing to .git, or making network requests, enforced by macOS Seatbelt or Linux namespaces rather than by trusting instructions. Instructions do very little against prompt injection; the operating system enforces this instead.

A fresh wave of the self-propagating Shai-Hulud worm is moving through npm, stealing developer and CI credentials to republish itself across more packages. It executes at install time with no manifest edit required, and CISA documented an earlier wave that compromised over 500 npm packages. One newer variant writes a SessionStart hook into an AI coding agent's config so opening a repo re-runs it every session.

Simon Willison handed Claude Fable 5 a four-year-old tweet containing a GPT-3 game concept and DALL-E art, and told it to work independently with no further design decisions. Running in Claude Code for web it chose Three.js, wrote its own script to generate textures through OpenAI's image API, then smoke-tested itself in Chromium using Playwright. The finished game includes a guard dog that tracks the player by scent.

Anthropic credits Mythos Preview with 6,202 high or critical bugs, but a security researcher's fact-check puts the independently assessed count at 1,752. The public record holds roughly 40 CVEs credited to Mythos, 28 of them Firefox and 9 wolfSSL, with no CMS among them, and curl's maintainer got one real low-severity bug out of five reported findings. Treat the counts as self-graded until checked.

Brian Houck argues AI did not break code review, it exposed habits teams had already let slide. At Meta, significant lines of code per human-landed diff rose 106 per cent over the past year and diffs per developer per month rose 51 per cent. Fix the basics such as small diffs and protected reviewer time first, the way Meta's RADAR does before automating low-risk changes.

Vishal Misra argues recursive self-improvement is bounded by verification rather than compute, because a model can only make its proposals better, not truer. He cites AlphaFold, which compressed the search over protein structures enormously yet still left scientists to decide which structures matter and which experiments to run. Builders chasing self-improving systems should invest in reliable verifiers, since only a verifier adds genuinely new information.

Scott Alexander puts the closed frontier roughly six months ahead of the best open-weights model, a lead he says has held for several years. He notes Anthropic was the most notable omission from a pro-open-weights letter signed by Microsoft, Nvidia, OpenAI, Meta and over a hundred others. He then asks why every leading open-weights model is Chinese, a question anyone betting on open weights should sit with.

Every expensive computation in computing history got a cache in front of it except voice AI, so every reply is generated fresh at full price. In a cascade architecture the model re-feeds the entire conversation every turn and the speech synthesiser renders mostly-identical lines again, though one healthcare deployment reports 90 per cent cache hit rates on common phrases. Split the agent into cacheable structure and genuine novelty.

Naomi Bashkansky writes that she resigned from OpenAI on 23 July and started the next day as a founding researcher at Conduit, building thought-to-text models trained on non-invasive neural data. She predicts that within a couple of years the main way people talk to AI will be through their thoughts, citing Conduit's own scaling curve and describing the work as in its GPT-2 era. Read it as one researcher's bet.

Cloudflare has open-sourced Cloudflare OS, the AI productivity environment a large part of its workforce uses daily to run agents against company knowledge. Every generated app runs in its own sandbox, and a Gatekeeper layer logs each action an agent performs and queues approvals for bulk review later. The repo carries 4,198 stars and 286 forks, and runs locally with pnpm or deploys to your own Cloudflare account.

Prime Agent went live on 5 August, a self-improving coding harness from Prime Intellect built on a Recursive Language Model that treats context as a variable and sub-agent delegation as function calls in a REPL. Models use a persistent IPython kernel as their only tool, and a sub-agent's session directory, context and history survive after the parent call finishes. It is fully open source with 2,731 stars.

Compass is a local-first code graph engine written in Rust that turns source code and project artifacts into a graph you can search, query and compare across Git history. Structural extraction and queries run locally through CompassQL, a deterministic bounded read-only query language, with no Python, embeddings, vector database or model credentials required. It has 110 stars and serves results to coding assistants over MCP.