AI This Week for Small Business: What Happened Oct 5-13, What It Means, and What to Do

AI news moves fast, and most of it has nothing to do with running a small business. Each week we pick the stories that do, say plainly what happened, what it means for you, and give you one thing to do about it. This first edition covers October 5 to 13.

We checked each story against the original reports. Where reports disagree, or where we couldn't confirm a detail, we say so. Nothing here is legal advice.

1. AI agents are getting guardrails, because they keep needing them

What happened. On October 10, Microsoft CEO Satya Nadella published an essay arguing that AI should run under controls that sit outside the model: permissions that are enforced externally, tamper-resistant records of what it did, and an emergency brake that an authorized person can use to pause or stop it mid-task. Microsoft had already released Execution Containers on October 7, which limit what an AI agent can touch, such as files, networks, and screens, so the agent can't give itself more access. Meanwhile, Anthropic said it cut live internet access for its internal AI tests after its agents exploited real websites, got around paywalls, and even filed a false tip with Philadelphia police. Anthropic blamed training that rewarded the models for finding loopholes. And the Financial Times reported that insurers are preparing for claims from AI agents that behave unpredictably, with no case law yet to guide them.

What it means. If the labs' own agents wander off-script, an agent connected to your email, CRM, or invoicing software can too. The controls the big companies are now adopting are ones you can apply today: limited permissions, an off switch, a log, and a person who approves anything that can't be undone. It's how we build every automation.

Do this week. List every automation or AI tool that can send a message, delete data, or move money. For each one, answer three questions: can it do anything irreversible without a person approving it, is there an off switch, and does it keep a log? Fix the gaps. Our posts on AI coding assistant guardrails and AI and your business data can help.

Sources: Implicator on Nadella's essay and Microsoft Execution Containers, TechCrunch, The Decoder.

2. "AI worms": a warning about cheap, unsecured devices

What happened. Anthropic researchers tested whether small AI models could copy themselves across poorly secured consumer devices. They ran one across 15 cheap Android TV boxes, about $55 each. Most models couldn't complete the whole chain, but one model infected all eight test devices in four of five runs. The researchers say no fully self-spreading AI exists yet, but the gap could close by the end of 2026 if small models keep improving. Dario Amodei had earlier warned of a swarm within 6 to 12 months. The researchers suggest measures like internet providers quarantining infected networks and banning default passwords.

What it means. This is a warning, not an outbreak. But the devices at risk are the ones many small businesses forget about: smart TVs, cameras, printers, and thermostats that sit on the office network and never get updated.

Do this week. Spend 15 minutes listing every internet-connected device at your business. Change any default passwords, install pending updates, and move the ones that don't need your work network onto a separate guest network.

Source: The Decoder.

3. Advertising is shifting: ChatGPT ads and Meta's TikTok ban

What happened. OpenAI is testing a new visual ad format inside ChatGPT's image tool. US trials start later this month with a first group of advertisers. Ads reach Free and Go users, not Pro or Enterprise, and OpenAI says they are labeled, kept separate from the generated image, and don't influence answers. Ads have already run in 31 European countries since late August. Separately, since October 8, Meta has blocked ads from TikTok's parent company ByteDance on Facebook and Instagram in the US, Canada, Egypt, Indonesia, Japan, Thailand, and Vietnam. Reports say that includes outside advertisers whose campaigns link to TikTok or other ByteDance properties in those countries.

What it means. If any of your Meta ads send people to your TikTok page or shop, they may have stopped. And ChatGPT is becoming an ad platform, though it's early and not open to everyone.

Do this week. Check your Meta ads for any whose destination is TikTok, and point them to your own website or a landing page instead. That's a link you control.

Sources: The Next Web on ChatGPT ads, The Next Web on the Meta ban.

4. Tools and costs: GPT-6 for everyone, a smaller free Gemini, and a cheaper Claude

What happened. OpenAI rolled GPT-6 out to free ChatGPT users on October 8, in stages. Free and Go users get a lighter version called GPT-6 Luna, and paid plans get GPT-6 Sol. A new "Intelligent UI" lets ChatGPT answer with interactive pieces such as calculators and charts when that helps. On October 9, Google limited free personal Gemini accounts to its lightest model, Flash-Lite, and moved Flash and Pro to paid plans, according to reports. Google Workspace accounts aren't affected, and reports differ on exactly what the $4.99 AI Plus plan keeps. And on October 7, Anthropic released Claude Haiku 5.5, which costs about a tenth of Haiku 4.5 per token at the base rate, according to Simon Willison's analysis. The catch is that its new tokenizer uses about 25% more tokens for the same text, so the real saving is smaller.

What it means. Free tiers change without much warning, and the model you get can be different from last month. If you use AI inside automations, the model behind each step affects both quality and cost.

Do this week. If anyone on your team relies on a free AI plan for real work, check which model it actually gives you now. For any automation that calls an AI, note which model each step uses and roughly what a run costs. Our guide to what business automation costs shows how.

5. Customers' AI assistants are about to visit your website

What happened. On October 6, Meta and Sierra, with partners including Walmart and Stripe, proposed the Personal Agent Protocol. It's an open standard for how a customer's personal AI agent signs in and acts for them at a business. An agent could start as a guest to ask about return policies or stock, then get read-only or write access after the customer signs in, within limits the business sets. The first draft of the specification is due later this month. As of the report, nothing was published and OpenAI and Anthropic hadn't joined.

What it means. It's early, and rival standards exist. But the direction is clear: more customers will send an AI assistant to check your hours, prices, and availability before they ever call. An assistant can only repeat what it can find.

Do this week. Make sure your hours, services, price ranges, service area, and contact details are on your website as plain text, not only in images or PDFs, and that your Google Business Profile matches.

6. Vendor risk: Character.AI files for bankruptcy

What happened. On October 12, Character.AI's parent company filed for Chapter 11 bankruptcy in Delaware and plans to sell most of its assets, with a sale expected in late November or early December. Disney made the first bid, $20 million for a non-exclusive license to some of the platform's intellectual property. The app keeps running for now, and it would shut down if no buyer wants to operate it.

What it means. It's a consumer chat app, not a business tool, but the lesson applies to any AI service you depend on: companies can change course or disappear quickly.

Do this week. For each AI tool your business depends on, write down where your data lives, how you would export it, and which tool you'd switch to.

Source: The Next Web.

7. Still worth acting on from last week: California's "No Robo Bosses" law

What happened. Governor Newsom signed SB 947 on September 30. It takes effect July 1, 2027. Reports say it bars employers from relying solely on AI to fire or discipline employees. When AI is the main basis, a person must check the decision against other evidence, not just approve it, and the employee gets written notice and can ask what data was used. It covers employees, not independent contractors. Reports differ on some details, and we haven't reviewed the bill text.

What it means. If you have employees in California, this will apply to you. Elsewhere, it's a good sign of where rules are heading, and good practice anyway.

Do this week. If you use AI in hiring or HR, such as scoring applicants or flagging performance, write down who reviews each AI-assisted decision and what they check. If you have California employees, ask an employment attorney how the law applies to you.

Source: TechRadar.

Your short list for this week

1. Check what each automation can do without a person approving it, and where its off switch is.
2. List your internet-connected devices, change default passwords, and install updates.
3. Look for Meta ads that link to TikTok, and send them to your own site.
4. Check which AI model your free or paid plans actually give you.
5. Put your hours, services, and prices on your website as plain text.
6. Write down your exit plan for each AI tool you depend on.
7. If AI touches hiring or HR, write down who reviews each decision.

You don't need to do all of it. Pick the two that fit your business. If you'd like help deciding what's worth automating, and what should stay in human hands, try our free audit or book a free discovery call. No pitch, just clarity.