TWP: Saving Private Data

Hi friend,

Three changes landed this week and they all do the same thing. They let AI act while you are not watching. Useful or unnerving, depending on how much access you hand over.

ChatGPT will now run jobs on a timer, including on the free plan. Scheduled tasks reached Free and Go accounts on Tuesday. Three active tasks, once a day at most. Plus accounts get five and can run them hourly (OpenAI). The more interesting part sits above that tier: on Plus, Pro and Business a task can now fire when something happens rather than when the clock ticks, triggered by a new Gmail message, a Slack post or a GitHub event. Tasks are shareable too, so a colleague can hand you theirs and you edit your own copy.

Set one up before deciding whether you like it. Give it a job you already do badly because it bores you.

"Every Friday at 4pm, search for news from the past seven days about [your industry body / your three biggest competitors / your council], and send me only items that would change a decision I make. If nothing qualifies, say so and stop."

That last sentence carries most of the weight. Without it, a recurring task invents relevance to justify itself, and you spend a month reading filler you asked for.

ChatGPT Work will now sign in to websites as you. Announced the same day: the work browser handles login screens and security codes, so it can get inside portals, forms and account pages instead of stopping at the door. Plus, Pro and Business, web and mobile (OpenAI). Passwords go through a secure prompt rather than into the chat, and anything with money or legal consequence asks you first.

The risk arrives when you pair that with the new inbox triggers. An agent watching your Gmail reads whatever lands, including an email written specifically to give it instructions, and OpenAI says plainly that prompt injection is unlikely to ever be fully solved. Aim it at an inbox you control, keep connected app permissions read-only where the option exists, and leave outbound actions off until you have watched it work for a few weeks.

Claude's memory became editable, which finally makes it worth reading. Anthropic extended memory across chat and its cloud workspace on Tuesday and put every stored item under Topics in Settings, where you can edit or delete any of it. A separate toggle controls whether health and belief topics get remembered at all. Memory runs on by default for Free, Pro and Max, and off by default for Team and Enterprise (Anthropic), so your work account and your personal one behave differently without saying so.

Open that list and read it. Most people find something they mentioned once, months ago, that they would not have chosen to keep.

Smaller things worth a line. Microsoft is rolling out Python inside Copilot's Excel editing, so it can build a working model rather than describe one, and Outlook emails now open inside Copilot Chat and Copilot Notebooks (Microsoft), all of which needs a Microsoft 365 Copilot licence on top of your normal subscription. Google is giving university students twelve months of AI Plus at no cost across 140-odd countries, New Zealand included, with a redeem-by date of 31 December and US$4.99 a month afterwards unless cancelled (Google). Gemini starts connecting to outside apps such as Otter and OpenTable over the coming weeks. And ChatGPT now loads long conversations faster and knows what time it is where you are, which sounds like nothing until you have asked it what is open tonight.

See you next week,

Shantanu

P.S. If someone forwarded this to you, you can subscribe at auracle.co.nz and get it yourself every Thursday. (Or set up a scheduled task to remind you to read it, which feels like exactly the sort of thing that ends badly.)