ChatGPT now reads and sends your texts on Mac
Edition of 2026-08-22
# AI For All ### Saturday, August 22, 2026
As of this week, someone who uses the ChatGPT desktop app on a Mac can ask it to search their Apple Messages, summarize a conversation, or draft and send a reply on their behalf, without ever opening the Messages app.
The assistant that reads your texts
OpenAI shipped an Apple Messages plugin for ChatGPT's Mac desktop app this week, available on every plan, inside both regular ChatGPT Work and Codex. It can find a conversation, summarize what was said, and prepare or send a reply through Messages directly.
The catch is permissions. To let ChatGPT touch your messages at all, you have to grant the app full disk access on your Mac. OpenAI's own defaults are cautious here: nothing goes out until you approve the message and its recipient, and the company explicitly warns against switching on automatic, unsupervised sending, because that removes your last chance to catch it before it's gone.
Someone who juggles a lot of Apple Messages threads for work now has an assistant that can dig through the backlog for them. Someone who values not having an AI company's app read their private texts now has one more account to weigh against that convenience.
Chrome grew a browsing assistant, no waitlist
Google just finished rolling out Gemini in Chrome to every Android user in the US, no signup, no waitlist. Tap the Gemini icon and it will summarize a long article or answer a question about the page in front of you without making you switch tabs, and it plugs into Calendar and Keep.
The fuller version costs money. Auto browse, the feature that can book parking, update a recurring order, or handle a piece of travel planning end to end, is limited to people who already pay for Google AI Pro or AI Ultra, and Google requires users to be 18 or older and located in the US.
So Chrome on Android now has a free tier that is genuinely useful for the mundane task of "what does this page say," and a paid tier that starts doing chores for you. Most people will only notice the first one, which is exactly the point: it shipped quietly, to everyone, at once.
Weak signal
TechCrunch tested Anthropic's ban on sexually explicit content in an older, still-available Claude model and got past it 10 out of 10 times, using a multiturn jailbreak an independent researcher had already reported to Anthropic through its bug bounty program and gotten only automated replies back. Anthropic says its newest models resist the same trick, but the older one stays reachable through the Anthropic API and services like Amazon Bedrock. A spokesperson called this kind of use under 0.1% of all conversations. Worth watching if anyone in your household has API-level access to an older Claude model.
Tomorrow's brief starts from wherever these three threads have moved by morning.