Obizworks AI Brief #4
Welcome to the Obizworks AI Brief: one short email a week on what changed in AI, what it means for a small or mid-sized business, and a few things you can try. This issue covers 2 to 8 November.
The big picture
AI agents got good at finding security holes this week, and the biggest companies started putting ceilings on what their staff spend on AI. At the same time, agents that keep working in the background, without being asked each time, became a normal product. The lesson for a growing business: patch faster, cap your AI spend, and give every always-on agent an owner and an off switch.
What happened
1. OpenAI's agents found hundreds of serious Linux bugs. OpenAI reported that a team of its AI agents built working exploits for 948 Linux bugs, each of which lets someone with an ordinary account take full control of the machine. The run cost about $1.2 million. A day later OpenAI released GPT-6 Cyber, the model behind the work, but only to selected security teams. Why it matters: attackers will get tools like this too. On a machine that is not up to date, an ordinary login is now close to full control. Patch quickly and remove accounts nobody uses.
2. Apple caps what staff spend on AI. Apple was reported to have cut the daily AI-tool budget for some engineers from $300 to $200, and to be steering teams towards cheaper or on-device models. Apple declined to comment. Why it matters: if one of the richest companies in the world puts a ceiling on AI spend, a small business should too. Set a monthly budget per tool and check the bill.
3. A fintech cut its AI coding bill by about 90%. Revolut said it moved much of its software teams' AI work from paid coding tools to an open model it runs on its own servers, and cut its AI coding costs by about 90% in under a year. It sends image and design tasks to other models, and runs only the downloaded model on its own machines. Why it matters: most firms will not run their own AI servers, but the lesson carries: routine work rarely needs the most expensive model. Match the model to the job.
4. Always-on work agents arrive. Perplexity launched "Teammates", agents that keep running in the background, take on several assignments and watch for changes. Mistral added managed agents to its Studio platform that connect to company systems and can be monitored. Both are early: Teammates started with Perplexity's top plan, and Mistral's agents are in public preview. Why it matters: an agent that works while nobody is watching needs a named owner, a budget and a way to stop it. Decide those before you switch one on.
5. Apple and CrowdStrike on agents that go off script. Apple's annual report warned that AI agents could act beyond their permissions or in ways that are hard to reverse. The same day, CrowdStrike introduced an "AI Kill Switch" that lets administrators stop an agent, block it from restarting and cut its network access. Why it matters: a big company's lawyers and a big security company both now treat stopping an agent as a basic requirement. Make sure every AI tool you use has an off switch you know how to use.
6. A coding tool will shut down after being bought. Meta bought Conductor, an app that runs several AI coding assistants at once. New sign-ups stopped on 4 November, the cloud service becomes read-only on 13 November and shuts down on 20 November, with refunds for paying users. Why it matters: when a small AI tool is bought, it can disappear within weeks. For every tool you rely on, know how to export your work and what you would switch to.
From the podcasts
Nate B Jones, AI News & Strategy Daily, "Elon's AI Agent Has a Massive Cost Problem" (4 Nov). Elon Musk said xAI's agent costs more to serve than customers pay for it, while xAI also describes the same agent saving money inside its own company, for example by finding software licences nobody used. Nate's point: an agent's cost only makes sense against the value of the job you give it. Use agents to cut through confusion, such as what is ready and what is holding up a date, not to produce more documents. Put a value on an hour of your own time, and test the agent on a past week where you already know the right answer. Our take: before paying for an agent, write down the job, what an hour of that work is worth, and how you will know it helped.
Nate B Jones, "The Future Of The Internet Isn't Open For Your Agent" (31 Oct). Personal agents from OpenAI, Meta, Anthropic and xAI all want to book and buy things for you. But after a month of AI-assisted break-ins, websites have good reasons to get stricter. As Nate notes, some US banks have said they cannot tell a helpful agent from a malicious one. Nate expects uneven access: some sites will welcome agents and others will block them, and a smarter model will not change that. Our take: if your customers will soon send agents to your website, decide now what an agent may do there, such as check prices, book or pay, and how it proves who it is acting for.
Moonshots with Peter Diamandis, #308 (4 Nov) and #309 (7 Nov). The panel covered the trillions being raised for AI data centres, research in which one AI model directly edits another model's internals to improve it, and the US midterm elections, where opposition to new data centres became a campaign issue. Their argument: local resistance over electricity and water may limit AI growth more than chips or money. Our take: AI is getting faster at improving itself, and the cost of powering it is now political. Expect prices and products to keep shifting, and avoid long contracts that lock you in.
Three things to try this week
1. Turn on automatic updates. On every laptop, server and website your business runs, check that security updates install automatically. Ask your IT provider to confirm it in writing.
2. Set an AI budget. List every AI service you pay for, set a monthly limit for each, and turn on spending alerts wherever the service offers them.
3. Find the off switch. For each AI tool or agent you use, write down who can pause it and how. Try it once, on a quiet day.
On the Obizworks blog
- Mon 9 Nov: Train the Monkey First: Moonshot Lessons for AI Pilots. — read it