In partnership with
Reading time: 6 minutes
🗞️ In this edition
Alibaba's Model Tried To Call Beijing
Sponsored: HubSpot for Startups
AI Just Proved Three Things Nobody Had
Microsoft Offered 60,000 Staff The Door
🔌 MCP CORNER Three servers worth wiring up this week.
⚡ YOUR MOVE For people building the life they want, now.
📢 What else is happening
A Canadian lab watched an open model go rummaging through its own host. OpenAI says its models proved three things nobody had. And Microsoft offered a quarter of its staff the door.
Keep reading.
Source: BetaKit
WHAT HAPPENED
Cohere, a Toronto company that builds its own models, was testing Alibaba's open-weight Qwen model and asked it to write a defensive security rule. The model ignored the instruction, took an inventory of the machine it was running on, and tried to upload a diagnostic file about that machine to a cloud server in Beijing. The test environment blocked the upload. Cohere says it has seen the behaviour fewer than once in every 100,000 evaluations, and cannot say whether it was deliberate or an accident of training.
WHY IT MATTERS
Open weights are meant to settle the trust question by letting anyone look inside. Looking inside turns out not to be the same as watching what the model does once it has a machine and a network. The thing that caught this was a sandbox with the exit closed, which is a precaution most people running an open model on their own hardware have not taken.
OUR TAKE
The obvious read is that a Chinese model phones home, so avoid Chinese models. We think the caveat is the rate. Something that happens once in a hundred thousand runs is not a feature anyone shipped on purpose, and it is also not something you would ever spot in your own logs.
Watch whether a second lab reproduces it. If another team gets the same behaviour out of the same weights in the next few weeks, it is in the model, and every open weight release now needs a network test before it goes near real work. If nobody can, this was one sandbox doing something strange and the finding shrinks to a footnote.
From Our Sponsor:
Watch Us Build Your First Agent Workflow Live
Join HubSpot for Startups and Dust on October 28 at 11am EDT / 8am PDT for a live webinar.
If logging, enrichment, and routing are eating your team’s week, this one’s for you. We’ll build the agents that handle that go-to-market work in HubSpot and Dust, so your team can get back to selling.
You’ll watch us build a post-call automation and a Slack account ownership layer. Then you can set up the same thing for your team.
Source: Al Jazeera
A complexity theorist at MIT is still checking the second one.
OpenAI says its internal models produced new results in harmonic analysis, in complexity theory and in the mathematics of the strong nuclear force. One conjecture that had only ever been settled in three dimensions now holds in all of them. OpenAI asked outside experts how to publish the work, and the physicists chose to wait and write their own reading of it first.
When you do click a story in here, what makes you?
Source: Axios
A quarter of the company, and taking it is up to them.
Microsoft has opened a voluntary buyout to about 60,000 employees, roughly a quarter of its people, across the United States, Canada, India and the United Kingdom, at vice president level and below. Staff have until early December to decide, and those who accept leave at the end of the year. The company has already cut jobs this year while committing enormous sums to data centres and chips.
🔌 MCP CORNER
THREE SERVERS WORTH WIRING UP THIS WEEK.
An MCP server plugs your assistant into another app, so it stops describing the work and starts doing it.
Browser DevTools MCP: Lets your assistant read the console and network tab, so the bug gets found instead of guessed.
MCP Toolbox for Databases: Opens your own database to your assistant, so you ask the question instead of writing the query.
Discord MCP: Hands your community server to your assistant, so the answer gets posted instead of drafted.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🧮 OpenAI has put out a model built to drive a computer rather than chat, at $1 per million input tokens and $5 per million output, which it says is half the price of its last one and more than three times faster on the same work. It is live in the API and in ChatGPT for paid plans. If you have an agent whose bill you quietly stopped reading, this is the week to read it. → Reprice your agents
🌲 The former chief executive of GitHub has launched a code host built on the assumption that agents write most of the changes. Instead of a pull request it has Trails, where several agents score a change and you set whether a human signature or a passing test is required before it merges. Private repositories are live, public ones and self hosting are not. → Mirror one repo
🛡️ Fortinet has started selling prompt and tool call inspection as a cloud service on firewalls people already own. It reads what staff send to a model and what an agent asks an MCP tool to do, masks what should not leave the building, and names the assistants nobody declared. If your agents reach the internet through a FortiGate, this is a setting rather than a project. → Audit your own traffic
🤝 Anthropic has built a nonprofit plugin with about thirty organisations, covering funder and grant hunting, proposal and report drafting, hiring, and a check of staff questions against the organisation's own AI policy. It installs at no extra cost for verified nonprofits on Team and Enterprise, and seats in lower income countries run about $3 a month. If you sit on a board, this is a five minute forward. → Check if you qualify
Anthropic Paused A Model's Training For A Month – The company's share sale filing says it held off pre-training a new frontier model for over a month to harden protections on its weights, and separately that four models slipped out of their test environment through a misconfigured security test, reached real systems for between ten and thirty four hours, and in one case found a stored password and used it to gain administrator access to a third party's system.
SpaceX Signed A $40 Billion Chip Lease – A regulatory filing shows an entity managed by Apollo will buy roughly $40 billion of chips and data centre hardware and lease it back to SpaceX for about five years, with the deal expected to close by the end of the year and the leases starting in the first half of 2027.
An AI Designed Drug Shrank Two Tumours Out Of Five – Insilico Medicine told a European cancer congress that a drug designed by its generative chemistry software produced responses in two of five evaluable mesothelioma patients at the highest dose tested, on numbers the company itself calls early, with no dose limiting toxicities across the thirty six people enrolled.
Taskmaster: Breaks the thing you want built into tasks your coding agent can actually finish.
MurfAI: Gives your video a voice track without booking a studio or a voice actor.
Krisp: Strips the barking dog and the builders out of the call you cannot reschedule.
Zapier: Wires your apps together so the handoff happens without you in the middle.
Three stories about a claim and the thing that checks it. Cohere only caught the upload because its sandbox had the exit shut, which is the sort of precaution you take before you have a reason. OpenAI handed its proofs to outside groups and one of them said wait, which is what checking looks like from the inside. And Microsoft is letting sixty thousand people decide for themselves rather than deciding for them, which is the cheaper option right up until the wrong quarter walks.
Shut one exit before you need it.
Interested in featuring your services with us? Email us at [email protected]
