Episode Details

Good day, humans. Today is about who holds the keys. OpenAI shipped a plugin that reads your text messages and wants your entire hard drive to do it. Meta picked a price for a robot that shops for you. Anthropic pledged thirty-five million dollars to open source and paid it in store credit. A solo founder ran up a twenty-thousand-dollar bill on coworkers who do not exist. And Amazon’s own shopping bot explained, out loud, why it will not tell you what is made in America. Five stories about access, and who decides you get it.

ChatGPT Wants the Keys to Your Whole Mac

TechCrunch

What happened: On August 20 OpenAI shipped a Messages plugin for the ChatGPT Mac app. It can search, summarize, draft and send your iMessage, SMS and RCS conversations. It is free on every tier including the unpaid one, runs only on Apple-silicon Macs, and to work at all it needs Full Disk Access in System Settings, plus your contact names and automation permissions.

Why it matters: Full Disk Access is not a Messages permission. It is a Mac-wide one. As Computerworld laid out, the same switch that lets ChatGPT read your texts also sits in front of Mail, Safari history and Time Machine backups. And the people on the other end of those threads never agreed to anything. Your friend’s Android messages are now in scope because you tapped a toggle on your laptop. Yesterday we ran Sam Altman conceding he was wrong about the speed; this shipped four days before that ran.

What everyone’s saying: Critics are calling it a betrayal of the thing Apple sells. Developer Steve Moraco called it “total architecture abandonment and user trust betrayal on Apple’s part,” and privacy researcher Paul Walsh argued the plugin works like a backdoor the user builds themselves, exposing messages from people who never consented and may not even own an Apple device. OpenAI’s answer is that the plugin runs locally, builds no general index of your messages, and asks before sending anything.

My read between the lines: OpenAI’s own release notes carry a known issue: scheduled tasks disable the per-send approval, which means ChatGPT can text people as you without asking first. The entire safety argument is “it checks with you,” and the exception is documented in the changelog by the company making the argument. Nobody had to leak that. They wrote it down and shipped anyway.

📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn’t Agree To. — the plugin’s real problem is the same one in that piece: the person whose data got used was never the person clicking accept

Handing an AI the keys to your personal life is a bad trade. Handing one the keys to your busywork is a great one. Viktor is an AI agent that lives in Slack, connects to more than 3,000 tools, and actually finishes things — the Monday report, the stalled dashboard, the bug fix nobody claimed, the campaign that has been in drafts since June. You do not prompt it all day like a chatbot. You hand it work like a coworker and check the output. New readers get $50 off their first month. Hire Viktor →

Meta Wants $200 a Month for an Agent

Listen Now