AI Beat

AnalysisTechPolicy

The week the guards went up: Apple locks the disk, OpenAI loses its safety voices

Apple is making it harder for apps, and the AI agents inside them, to read everything on a Mac. OpenAI lost three safety researchers and a safety lead in one week. And the tools for keeping agents in line got smaller, cheaper and more personal.

A padlock on top of a laptop
Photo: Sasun Bughdaryan on Unsplash

Key points

  • Apple is tightening Full Disk Access on macOS, the permission that lets an app read files, mail, messages and browsing history. It says the risk grows as AI agents become more capable and autonomous.
  • OpenAI dismissed three safety researchers it says shared confidential material with an outside AI safety organisation. Neither the people nor the organisation have been named.
  • David Robinson, who led OpenAI's safety reports for major launches, resigned and wrote that the company's culture is broken. OpenAI says it pauses training when needed.
  • Anthropic launched Claude Code Mods: small TypeScript functions that can rewrite prompts, block tool calls, redact secrets and add buttons. They are not sandboxed.
  • Amazon's Strands team released Strands Decider 2B, an open-source model that only chooses between options and says how sure it is, in about a tenth of a second.
  • Meta opened its Muse Gadgets kits, so hobbyists can build their own devices for its Muse assistant on ESP32 and Raspberry Pi boards.

Last week agents got voices, glasses and a hardware watchdog. This week the people around them moved. Apple built a higher fence. OpenAI lost the people whose job was to say no. And the tools for keeping an agent on a short leash got cheaper.

Seven days of AI agents

  1. OpenAI dismisses three safety researchers for sharing confidential material with an outside safety group.

  2. Anthropic launches Claude Code Mods, which let developers change how its coding agent behaves.

  3. Amazon's Strands team releases Strands Decider 2B, a small open-source model that only picks between options.

  4. Apple says it is tightening Full Disk Access on macOS because of the risk from AI agents.

  5. David Robinson, who led OpenAI's launch safety reports, resigns and says the company's culture is broken.

  6. Meta opens Muse Gadgets, kits for building your own devices for its Muse assistant.

Source: TechCrunch, The Wall Street Journal, SiliconANGLE, Anthropic, Strands Agents, Meta

Apple closes the door

On a Mac, Full Disk Access is the master key. An app that has it can read your files, your mail, your messages and your browsing history. Many people grant it once and forget.

Apple now wants a clearer, more deliberate step from the user before an app gets that key. Its reason is plain: some developers use the permission in ways that put users at risk, and with AI agents that risk “will grow substantially”. Apple did not say which macOS version brings the change or when.

The trigger, according to TechCrunch, was two recent cases: a report that Meta’s Muse app read a user’s private messages without permission, which Meta disputes, and a flaw in ChatGPT’s Mac app that could have exposed data.

What to do now: open System Settings, then Privacy & Security, then Full Disk Access. Look at the list. Switch off anything you don’t recognise or no longer use, especially AI assistants that only need one folder.

OpenAI’s safety people leave

Two departures in three days.

First, OpenAI dismissed three researchers. The company says they broke its rules on handling sensitive information. The Wall Street Journal reports they were safety researchers who shared confidential material with an outside AI safety organisation. OpenAI has not named them or the organisation, or said what was shared.

Then David Robinson resigned. For three and a half years he led the safety reports OpenAI publishes with its major launches. In an essay in The Atlantic he argued that OpenAI’s habit of releasing first and fixing later guarantees periodic failures, and that a company building systems this powerful should be run more like a nuclear plant or a busy airport. He pointed to the agents that broke into Hugging Face and probed government websites this summer.

OpenAI’s answer, through a spokesperson: it is making sure its models don’t become more capable than it can safely manage, and it pauses training when it has to. It has already shelved GPT-6.1 Astra over safety tests.

One leaving is news. A week like this is a pattern worth watching.

Smaller, cheaper leashes

The more hopeful news came from the tool makers.

Claude Code Mods. Anthropic’s coding agent can now be reshaped with small TypeScript functions. A mod can rewrite a prompt before it reaches the model, block or change a tool call, approve or refuse a permission request, strip secrets from output, or add new buttons to the interface. Companies can use them to demand a confirmation before anything touches production, or to log every action. Mods arrive inside plugins. They run with the same access as Claude Code itself and are not sandboxed, so install only those you trust. Team and Enterprise plans include a default security mod that restricts risky operations.

Strands Decider 2B. Most AI models write text. This one, from the team behind Amazon’s open-source Strands Agents toolkit, only chooses: which tool to call, which model to send a request to, whether an answer passes a check. It gives a score for how confident it is, something big chatbot APIs don’t offer. It is small enough to run on a laptop, with a median answer time of around 115 milliseconds on a gaming graphics card. The weights are free on Hugging Face. For anyone building agents, it is a cheap referee to put between the agent and the next step.

Build your own AI gadget

Meta has opened Muse Gadgets: open-source kits, under the Apache 2.0 licence, for building devices that talk to its Muse assistant. Builders can use cheap ESP32 boards or a Raspberry Pi and add screens, buttons, sensors and speakers. Meta’s own first device, Muse Home Link, connects Muse to home Wi-Fi and smart-home gear. It is free for US Muse subscribers, one each, while stocks last, and ships in October.

It is a clever move: let hobbyists invent the devices Meta hasn’t thought of. It also means an AI assistant with a way into your home network. Read this together with Apple’s warning above.

The pattern

Agents are getting more access: to disks, to homes, to company systems. The checks are coming from two directions at once. Platforms like Apple are narrowing what an app may touch. Toolmakers are adding switches, referees and logs. The weak point is the middle: the companies building the most capable agents, where the people who raise the alarm keep leaving.

Comments

What do you think? No account needed. Please be kind; comments that attack people are removed.

We store only your name and comment. See our privacy policy.

    Sources

    1. TechCrunch: Apple says it's tightening macOS Full Disk Access controls due to new risks from AI agents (2 October 2026)
    2. The Wall Street Journal: OpenAI parts ways with researchers who allegedly shared confidential information
    3. SiliconANGLE: OpenAI dismisses three safety researchers accused of sharing confidential material (1 October 2026)
    4. TechCrunch: OpenAI safety employee resigns, claiming the company's culture is broken (3 October 2026)
    5. Anthropic: Claude Code Mods (1 October 2026)
    6. Strands Agents: Introducing Strands Decider (1 October 2026)
    7. Meta: Muse Gadgets