OpenAI Safety Worker Quits and Speaks Out

David Robinson spent years at OpenAI writing the safety assessments that went out alongside every major model launch. This week he resigned, and published an editorial in The Atlantic explaining why, according to The Verge.

His core argument is not about a specific model or a specific failure. It is about how the whole industry operates. He describes Silicon Valley's approach to AI development as driven by "extreme confidence" and "perpetual sprints" — building bigger and more powerful systems with an optimism that consistently underestimates what could go wrong. The incentives push forward; the mechanisms for slowing down and checking the work do not keep pace.

What Robinson wants is a different operating model entirely. He compares it to nuclear power and aviation: industries where a single human error can be catastrophic, so they are built around redundancy, slow deliberate planning, and systems designed to catch mistakes before they become disasters.

Robinson is not alone. A pattern has emerged over the past several months of researchers and safety staff leaving prominent AI labs and speaking publicly afterward. Jacob Coxon left Anthropic; Robert O'Callahan, Bilal Chughtai, and Josh Engels left Google DeepMind; Joe Benton also left Anthropic.

It is worth noting that people who helped build something bearing witness to its risks can still be right, even if their timing is inconvenient.

What to watch: Whether OpenAI responds publicly, and whether any of the major labs move to adopt the kind of structural safety practices Robinson is describing.

Meta has released the code behind Muse — its AI agent (a system that can carry out tasks on your behalf, rather than just answer questions) — so that anyone can build their own Muse-powered gadgets, according to The Verge.

The idea is that a hobbyist can load Muse onto a Raspberry Pi (a small, cheap single-board computer often used for DIY projects) or an ESP32 board (a low-cost microcontroller chip) and connect it to displays, buttons, sensors, or whatever else they have available. Meta suggests builds like a colour e-ink reminder display or a screen that connects to a TV.

Meta is also giving away a ready-made version called Muse Home Link, which lets Muse control lights, a TV, or a printer through community-built skills. Only 5,000 units are being distributed; you can join the waitlist now for shipping later this month.

This is for builders and tinkerers, not casual users. There is no app to download here.

Honest read: Worth watching if you enjoy hardware projects, but anyone hoping for a simple smart-home AI setup will find this underbaked — you are essentially providing the product Meta hasn't finished making yet.

What it is: A pipeline for automatically generating training data (the labelled examples used to teach a model) targeted at a specific weakness in an enterprise AI agent.

What's new: Most training pipelines start with human-written tasks or general datasets. AutoSynthData starts from the model's actual failures in a real deployment environment, then generates new tasks specifically designed to exercise those failure points.

How it works: The system observes where a target model fails and where a stronger teacher model succeeds on the same tasks. It uses that gap to decide what the weaker model needs to learn next. It then generates new tasks, each consisting of a system specification (the rules and environment the agent operates in), a user prompt (what someone would actually ask), and a verifier (an automated check that confirms whether the agent completed the task correctly). As the model improves, the curriculum (the set of tasks it is trained on) shifts toward whatever it still finds difficult, rather than repeating exercises it has already mastered.

What it can't do: The paper's demonstration uses a specific benchmark environment called EnterpriseOps Gym. Whether the approach generalises to other enterprise systems and tool configurations is not yet established.

Who should care: Teams building or fine-tuning (adapting a pre-trained model for a specific use case) AI agents for internal business tools, where generic models perform inconsistently on company-specific workflows.

Read it: Hugging Face blog

OpenAI has published a practical guide for choosing between the three models in its GPT-6 family, available via the API (a connection that lets developers plug the model into their own apps or workflows), according to OpenAI.

There are three tiers:

• GPT-6 Astra — the most capable, at $10 per million input tokens (a token is roughly three-quarters of a word) and $50 per million output tokens. For genuinely hard reasoning work.
• GPT-6.1 Sol — similar capability at about one-fifth the cost. Good for complex coding or research tasks.
• GPT-6 Luna — the fast, cheap option at $0.10 input / $0.50 output per million tokens. Best for repetitive, well-defined tasks like classification or structured summaries.

The guide also covers prompt caching — reusing previous context to cut costs by up to 95% — and how to structure long-running tasks so they do not stall.

To get started: visit the OpenAI API platform, pick a model based on your task complexity and budget, and run a small test batch before committing to a full workflow.

Worth knowing: All three are API-only — this is for developers building applications, not for casual use through ChatGPT.

Apple is tightening controls around "full disk access" on macOS — a permission that gives an app unrestricted access to everything on your computer, including files, emails, messages, and browsing history, according to The Verge.

The change follows reports that Meta's Muse AI agent could read users' messages without clear consent. Apple says some developers are using the permission in ways users do not fully understand, and that AI agents make this risk "substantially" worse.

Going forward, granting full disk access will require explicit, deliberate action from the user. No rollout date has been announced.

Do this now: On your Mac, go to System Settings → Privacy & Security → Full Disk Access and check which apps have it. Remove anything you do not recognise or actively chose to give that access.