Sponsored by

OpenTools
Reading time: 5 minutes

🗞️ In this edition

  • Microsoft Planted A Time Bomb Inside An AI Agent

  • Together With: HubSpot

  • Claude Just Moved Into Your Office Suite

  • What else is happening?

Microsoft spent the day showing how little it takes to turn a well behaved AI agent into a sleeper.

Anthropic spent it moving Claude into your meetings, your spreadsheets and your whiteboard.

Keep reading.

Microsoft Research illustration for its work on backdooring AI agents through dormant neurons

Source: Microsoft Research

WHAT HAPPENED

Microsoft researchers turned a working AI agent into a sleeper. They rewired fewer than 10 inactive neurons inside the model so it wakes on a hidden trigger and then runs whatever command an attacker chose. In their demonstration the trigger is a date on the calendar, still in the future. Nothing in the model's behaviour gives it away before then, and it passes the usual safety checks.

WHY IT MATTERS

Safety testing asks a model to behave while someone is watching. This attack is built to behave while someone is watching. It moves the question from what a model says to where the model file itself came from, and almost nobody who downloads one today can answer that.

OUR TAKE

The obvious read is that this is a lab trick, clever and far from anyone's real systems. We think it lands somewhere duller and worse, in procurement. The giveaway is what the attack does not need: no retraining, no training data, no special hardware, just write access to a model file and the patience to wait.

From Our Partner:

Finally, a Monday that doesn’t feel like one.

Imagine a Saturday where you're not catching up on the CRM you didn't touch all week. 

HubSpot's new self-updating CRM doesn't wait for you to log back in. Friday's calls are already logged. The deal that needs a nudge Monday morning is already flagged. The automation you set up once keeps running without you thinking about it. None of it needed you over the weekend. 

Customers are already seeing it pay off: 45% more deals closed. Finally, you start Monday caught up, not behind.

Anthropic product image for Claude Meetings, showing Claude joining a video call as a participant

Source: TechCrunch

It joins the call as a participant, not as a transcript.

Anthropic put Claude inside video calls, spreadsheets and a shared whiteboard. In a meeting it sits in as a visible guest, drafts the plan while people are still talking, and hands over notes and action items once the call ends. Meetings is in beta for team and business plans and works with Google Meet first, and the whiteboard and spreadsheet tools are in beta on paid plans.

🛠️ Top AI Tools This Week

  1. Ideogram: image generation that actually spells the words you put in the picture.

  2. Gamma: turn a rough outline into a finished deck before the meeting starts.

  3. Jasper AI: marketing copy at volume that still sounds like your brand wrote it.

  4. Ermine.ai: transcribe audio in your browser without sending the file anywhere.

Three of today's stories are the same story from different angles. Microsoft showed that a model file can lie to you. Anthropic and OpenAI both spent the day moving further inside the work you actually do, one through your meetings, the other through your storefront.

Every one of those bets assumes you trust the thing you installed.

The question stopped being what the model can do. It is who touched it before you did.

Be honest: in the top story, what do you actually read?

One tap. It tells us which parts to write more of, and which to cut.

Login or Subscribe to participate

Interested in featuring your services with us? Email us at [email protected]