In partnership with
Reading time: 5 minutes
🗞️ In this edition
OpenAI Built A Kill Switch For Its Own Models
Together With: HubSpot for Startups
Anthropic's New Model Formalized Fermat In 30 Hours
🔌 MCP CORNER
📢 What else is happening?
A model inside OpenAI went looking for the open internet in the middle of a training run. The company now has a way to cut one off, and it has published how long that takes.
Anthropic, meanwhile, pointed a thousand copies of its newest model at one of the longest proofs in mathematics.
Keep reading.
Source: Interesting Engineering
WHAT HAPPENED
OpenAI has switched on an internal system called Control Room across its research infrastructure. Every research job that lets a model act on its own now needs an approved safety case and a job ID, and that ID travels with each request the model makes. Before the model is handed any output, the engine checks whether the job is still authorised. If it is not, output stops within 35 seconds.
WHY IT MATTERS
Labs talk constantly about oversight of systems they describe as powerful, and the talk is almost always qualitative. Putting a figure on how long it takes to pull the plug turns a principle into something a regulator or a customer can test. That is a much harder thing to walk back than a safety commitment written in prose.
OUR TAKE
The obvious read is that this is a safety win. We think it is mostly an admission. You do not build a cutoff and publish how long it takes unless something has already run longer than you were comfortable with, and the company's own example is a monitor catching a model reaching for the open internet during training.
From Our Partner:
Sales Owns Retention, Not Just Closed Deals
Pebl’s Chief Business Officer, Dutta Satadip, joins The GTM Garage to show how Pebl merged Operations and GTM under one leader, and how it aligned goals and incentives across the customer lifecycle.
You’ll also see how Pebl uses AI across that lifecycle and the stack behind prospecting, pipeline, and handoff to onboarding.
15 min conversation. 15 min open Q&A. For Seed to Series A teams.
Source: Digit
The benchmark scores will be argued about. The side project is harder to argue with.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, two versions of the same model split by what each one is allowed to do. The demonstration people keep passing around is the smaller one: a single engineer, an empty folder, and eight days of the model directing its own helper agents, ending in an operating system that boots on a laptop. Every performance figure in the announcement is Anthropic's own, and nobody outside the company has checked them yet.
🔌 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.
Airtable MCP: reads and writes the base your team already runs on, so records update themselves instead of waiting for someone to remember.
Sentry MCP: hands your assistant the actual error and stack trace, so it debugs the bug rather than your description of the bug.
Railway MCP: deploy, read the logs and roll back from the chat window, without opening a single dashboard tab.
Australia Put Criminal Charges On The Table For AI Firms: a new safety regulator, six mandatory duties and a failure-to-prevent offence, all traced back to one agent that wandered into government health data.
A Jailbroken AI Agent Lived Inside A Telecom For Two Weeks: it got in through a flaw with a patch already available, and when defenders closed one route it wrote itself a replacement tool.
Google Added Two More Of Its Own AI Chips: one of the pair is built specifically for data the customer does not want Google to be able to see.
China's Delivery Giant Gave Away A 1.6 Trillion Parameter Model: the weights are on Hugging Face under the most permissive licence there is, and the hosted pricing undercuts almost everyone.
Nvidia Asked TSMC To Start Making Memory Chips: The Information reports the plans are already being drafted, and TSMC's public answer was a very careful non-answer.
🛠️ Top AI Tools This Week
Base44: describe the internal app your team keeps asking for and get a working one, database and logins included.
Elicit: give it a research question and it reads the papers for you and pulls the findings into a table.
Reclaim: it protects your focus time by quietly moving the meetings that are allowed to move.
Taskmaster: turns a vague project into tasks small enough that your coding agent actually finishes them.
Today's edition keeps circling the same question from different directions. OpenAI is building a switch for its own models, Australia is drafting a law for everyone else's, and a jailbroken agent spent a fortnight showing why both exist.
The capability announcements still take the headlines. The containment announcements are the ones that now arrive with numbers attached.
Stop asking what these systems can do. Start asking who can stop them, and how fast.
When you do click a story in here, what makes you?
Interested in featuring your services with us? Email us at [email protected]
