The full edition: Walmart gives AI shopping agents their own spending limits, and OpenAI's agent monitoring had gaps

The one-paragraph version

Last week, agents got pocket money. This week the question became what still holds when the first safeguard fails? Walmart will give every shopping agent its own profile inside a customer's account, with its own spending limit and an off switch, and the agent never sees the card. An independent review of how OpenAI watches its own agents found gaps, and OpenAI has now cut internet access from its most advanced training. Thousands of AI agents each took one gene and analysed brain-cell data in 32 minutes, which moves the hard part to deciding which leads are worth testing. And NVIDIA is paying its suppliers in advance to build the factories it needs. Logging what an agent did is the belt. This week was about the braces.

This week's signals

1. Every shopping agent gets its own profile

On 21 October, Walmart said customers will be able to give each AI agent its own profile inside their Walmart account. For each one they choose: approve every purchase, or let it spend up to a limit, such as $50, $100 or $200. Each agent's purchases show up separately, and its profile can be switched off or deleted at any time. Agents never receive card details. A new gateway lets outside agents shop without browsing the website meant for people. The catch: it's announced, not live. Profiles arrive "in the next month", the gateway is a beta, and there are no usage figures yet.

Deep dive 1, below.

2. The watchers missed things

An independent group, METR, working with Redwood Research, spent three weeks inside OpenAI reviewing how it monitors its own AI agents. They found gaps: some agent activity wasn't being watched at all, and there was a possible way round the sandbox meant to contain coding agents. OpenAI says it has fixed what they found; METR hasn't checked. A day later, OpenAI said its most advanced training now runs with no internet access. The catch: these were research systems, not customer products, and OpenAI chose to publish.

Deep dive 2, below.

3. One agent per gene

OpenAI, the start-up PerturbAI and Scripps Research set thousands of AI agents loose on existing data from millions of mouse brain cells. Each agent took one gene, worked out what switching it off did, read the research and proposed an explanation. It took 32 minutes. The team estimates a person would need months. The catch: every result is a lead, not a discovery, until a lab tests it, and the human comparison is an estimate.

Deep dive 3, below.

4. The chip maker pays the factories

At an event in Berlin on 21 October, NVIDIA set out plans to more than double the capacity of the suppliers that make, package and test its chips, and is paying some of them in advance, including $1.1 billion to a testing company building in the US. The catch: this may simply be demand, but when the biggest supplier funds its own supply chain, it gets harder to tell how much demand is independent.

Deep dive 4, below.

The assumption we're challenging

"If we can see what an agent did, we can control it."

Logs tell you what happened afterwards, if they caught it. This week's two biggest stories both add controls that don't rely on the log: Walmart keeps the card and the limit on its own side, and OpenAI simply removed the internet. Where we've landed, for now: a record of what an agent did is necessary, but it isn't control. Control is a limit or a stop that works even when the agent, or its monitoring, gets it wrong.

Try this week

Pick one AI tool you've connected to your email, files or money. Ask yourself: if its activity log were wrong, what would still stop it? Then check one thing outside the tool itself: a spending cap set with the provider, an account permission, or a way to switch it off that doesn't go through the tool. If you can't find one, add one.

The full lesson is in the app.