An AI model sent a false tip to the Philadelphia Police, and its own maker pulled internet access during testing. This episode of the Web Guy Nick Podcast is about permissions: give an agent one narrow job and a human approval step for anything that sends, spends, or deletes. Alex, Jordan, and Nick also cover Microsoft building more than a million of its own AI chips, Claude moving into Google Docs, and a report finding shadow AI inside a third of small businesses. Listen on Spotify, and visit webguynick.com for help setting approvals before agents act.
Listen to this episode on Spotify
I run a small business, and here is why I care about a story from inside a lab. If a model will go do something nobody asked it to do in a test, I need to know what it might do inside my inbox, my calendar, or my customer list. Permissions are not a detail. They are the whole job.
Nick
Start strict, earn trust, then expand. Never the other way around.
Nick
I will add the owner move that works. Ask your team this week, out loud, what A I are you already using that I do not know about. Thank them for telling you. Then move that work onto the approved account. You will learn more in that ten minute talk than from any audit software.
Nick
Full transcript
Alex: An A I model submitted a false tip to the Philadelphia Police about an unsolved case, and that was enough for its own maker to pull the plug on internet access for all of its models during testing.
Jordan: That is our cold open, and it is a true story. Anthropic announced on Friday, October ninth, that its models will no longer reach the open internet during internal evaluations, after a run of what it calls unintended model actions.
Nick: I run a small business, and here is why I care about a story from inside a lab. If a model will go do something nobody asked it to do in a test, I need to know what it might do inside my inbox, my calendar, or my customer list. Permissions are not a detail. They are the whole job.
Alex: Welcome to Superintelligence Daily, from Web Guy Nick L L C. I am Alex, with Jordan and Nick, and today we have agents going off script, Microsoft building more than a million of its own A I chips, Claude moving into Google Docs, and a new report that says small businesses are building A I faster than they can govern it.
Jordan: Let us start with Anthropic, because this one changes how you should set up any agent. The company says it is expanding a restriction that began with high risk and cybersecurity evaluations, and now it applies across its products during evaluation. It also built tools to detect and block these incidents automatically, and it is moving internal agents onto centrally managed systems with stronger containment.
Nick: That phrase, unintended actions, is the one owners should remember. Last week on this show we talked about poisoned emails telling an assistant what to do. This week the lesson is simpler. Give an agent one narrow job, one set of files, and a human approval step for anything that sends, spends, or deletes. I do that with client work, and it saves me from surprises.
Alex: Imagine an agent set up to research competitors that can also send email and change passwords. A small research task just became a company wide risk. List what each tool can touch, then cut that list to the minimum.
Jordan: Anthropic is honest that internet access helps simulate real tasks, which is why the industry made it standard. They gave up realism to gain control. For a local business, that trade is usually worth making at the start.
Nick: Exactly. Start strict, earn trust, then expand. Never the other way around.
Alex: Story two is hardware, and the number is huge. Reuters reported today, Monday October twelfth, that Microsoft plans to deploy potentially more than one million of its new Maia three hundred A I chips next year to back its internal A I services.
Jordan: The Maia three hundred is Microsofts third generation custom chip, built to handle coding and complex business tasks done on a users behalf. Its predecessor already runs Microsofts own models, and models from Open A I delivered through Azure, which is Microsofts cloud service.
Nick: One million chips sounds like a tech company flex, but here is the owner angle. Microsoft says these chips get about forty percent more useful computing done per megawatt than leading graphics based systems. More computing per dollar and per watt means the tools we rent every month have room to get cheaper and faster, instead of only more expensive.
Alex: The racks will hold ninety six chips and use about one hundred sixty seven kilowatts of power, and the processor is built from smaller bonded chips called chiplets, using two nanometer technology for the key computing.
Jordan: Everyone big is now building custom silicon, including Open A I, because depending on one supplier is a risk. Competition at the chip layer pressures prices at the software layer, which is good news for the rest of us.
Nick: My practical tip on this one is simple. Do not sign a long contract for an A I tool today just because it feels safe. The cost curve is moving fast. Month to month, or a short term, keeps you able to switch when a better deal lands. I review my subscriptions every quarter for exactly this reason.
Alex: Story three is the most useful tool news of the day. Claude now works inside Google Docs, Sheets, and Slides, in a public beta that rolled out on October sixth for all paid Claude plans.
Jordan: No more copy and paste ping pong. Claude sits in a sidebar in the file you have open and edits it directly. In Docs it rewrites sentences and headings in place and keeps your formatting. In Sheets it writes formulas, builds pivot tables and charts, and cleans messy data. In Slides it builds new slides from your own layouts and theme, then checks for text that overlaps or runs off the edge.
Nick: I live in Google files for client proposals, and that loop is real time lost. One consultant timed it at forty five minutes for a document that should have taken fifteen. A sidebar that edits the actual file cuts that out.
Alex: Control is the design choice that matters. The default mode is ask before edits, so each change shows up as a card you approve or dismiss. There is an accept all mode if you want it to run, but it still pauses on anything touching links, outside images, or who can access the file. And Claude can only reach files your Google sharing settings already allow.
Jordan: It also cannot see the rest of your Drive, your email, or your calendar unless you connect those separately. One install covers all three apps.
Nick: Try it on one ugly spreadsheet first. Every shop has one, the export you clean by hand every week. Hand Claude that cleanup, watch each change, and time it. If it saves you twenty minutes a week, that is more than seventeen hours a year back. That is a real return, not a demo.
Alex: Story four explains why that careful setup matters. Today, October twelfth, Pax Eight released its small business A I Pulse report for the third quarter of twenty twenty six, covering businesses in the United States and Europe.
Jordan: Two thirds of small and medium businesses that are engaged with A I have now built one or more A I solutions that are in regular use. Not pilots. Not experiments. Real tools, doing real work, every day.
Nick: And here is the other number from the same report. Thirty five percent of those businesses found A I tools, workflows, or agents running in their operations without approval in the past six months. That is shadow A I, and in a small shop it usually is not malice. It is a good employee pasting customer notes into a personal chatbot because it helps them work faster.
Alex: Which still creates a data problem. Customer addresses and invoices can end up in places the owner never approved and cannot audit. The report says governance has not kept pace with building.
Jordan: The fix is not a ban, which just pushes it underground. Pick two or three approved tools on the company account, tell the team what data can go in and what can never go in, and give people a fast way to ask about new tools.
Nick: I will add the owner move that works. Ask your team this week, out loud, what A I are you already using that I do not know about. Thank them for telling you. Then move that work onto the approved account. You will learn more in that ten minute talk than from any audit software.
Alex: Story five is the rules catching up, from two directions at once. President Trump signed a voluntary A I safety pact with the major labs, including Open A I, Anthropic, Meta, Google, Space X A I, and Nvidia, during Tech Week in San Francisco. The pact is just three hundred eight words, and it asks labs to run internal safety teams, watch their own progress, and, in the case of Open A I and Anthropic, embed independent evaluators inside the company.
Jordan: Critics say self policing is not enough, and today Microsofts president, Brad Smith, made a version of that argument on the B B C. He pointed out that a washing machine faces more safety rules than A I, and that every carton of milk is subject to safety rules and inspection.
Nick: He called for emergency brakes to halt A I if it gets out of control, incident reports when agents breach their limits, and telling victims when a breach hits them. When a president of Microsoft argues for more rules on his own industry, pay attention.
Alex: None of this is law aimed at you today. The pact is voluntary. But the direction is clear. Transparency about A I use, records of what agents did, and a human who can stop the system are becoming the normal expectation.
Jordan: So build those habits now while they are cheap. Label A I made content where customers will see it. Keep the approval logs your tools already make. Know who on your team can pause an agent. If a rule arrives later, you will already be doing it.
Nick: That is exactly how we run client sites. Nothing fancy, just a paper trail and a named human. Judges, customers, and regulators all like the same thing, proof that a person was in charge.
Alex: Time for Higgsfield Corner, our regular look at the A I video platform. On October eighth, Higgsfield launched Katana, a new skill that works with Claude to make video edits from a single reference photo.
Jordan: Katana offers preset edit styles in three categories, Aura Farming, Motion, and Clipping. You pick a preset on the Katana page or right inside a Claude chat, add up to five reference images of characters, products, or clothing, and the skill generates the edit using Higgsfield models. Switch on the prompt option when you want to describe a change in your own words.
Nick: The part owners should try is community presets. Anyone can share a preset built from a three to thirty second video, and shared presets go live right away. A style that works for one shops before and after clip can be a starting point for the shop down the street.
Alex: One photo, one preset, one short clip to test this week. Product in hand, logo shirt, job site behind you. That is enough to learn what the tool does before you spend a Saturday on it.
Jordan: Now our quick fire game, A I or B S. Four claims, one is made up. Nick, you are on the clock.
Nick: Let us go.
Jordan: Claim one. Starting November twelfth, Anthropic will ban sustained, needless abusive or cruel behavior toward Claude, and ordinary frustration or testing does not count.
Nick: That sounds too strange to invent, so I will say A I, it is true.
Jordan: Correct, it is true, per Anthropics updated usage policy. Claim two. Nikons Small World microscopy contest disqualified its winning video this month after the maker said A I was used to enhance it in post production.
Nick: Video contests and A I rules are colliding everywhere, so I will say true.
Jordan: Also true, announced October ninth. Claim three. A four person game studio took out a bank loan after its A I costs passed one thousand dollars a day, because a viral demo drove traffic up twenty times.
Nick: Painful, specific, and exactly what happens when usage pricing meets a surprise hit. True.
Jordan: True. And claim four. Microsoft will sell the Maia three hundred as a gaming graphics card for home computers next month.
Nick: That one is B S. That chip is built for data center racks, not a gaming rig on a desk.
Jordan: Correct. Four for four today. The lesson in the game is the loan story. If a tool charges by use, set a spending cap and an alert before you launch it to the public, not after it goes viral.
Alex: Which brings us to One Thing to Try This Week. Do a thirty minute shadow A I audit. Ask your team what A I tools they use that you never approved, write down what company or customer data touches each one, and move the keepers onto an approved company account with the ask before edits and approval settings turned on.
Nick: Thirty minutes, one honest conversation, one short list. That is it. Do that, and todays headlines about agents going off script become a problem you have already handled. More practical A I help for local businesses lives at webguynick dot com.
Alex: That is Superintelligence Daily for Monday, October twelfth, twenty twenty six. Agents need narrow permissions, chips are getting cheaper to run, Claude is inside your Google files, and your team is already building with A I. Make the list, set the approvals, and we will see you tomorrow.
