Daily AI brief · October 14, 2026
Gemini Will Now Work While You're Away, for $19.99
Google's paid Gemini plans can now run tasks while you're away from your computer, and Anthropic's researchers warn that cheap smart gadgets on an office network could soon host AI malware.
Google brings Gemini 4 Argon and "tasks" to the Gemini app (Google, Thurrott, Android Police) Google started rolling out Gemini 4 Argon, its newest top model, to Google AI Pro and Ultra subscribers in the Gemini app on Tuesday. Google says Argon answers more directly, with less hedging and repetition, and reports its progress more often on longer jobs. The bigger change for an office is a new switch in the main chat window that flips Gemini from chatting to doing: in tasks mode, Google says, Gemini can keep working for you when you're not at your computer, across your Google apps and more than 40 connected apps. Free users don't get either one, and Google's announcement names only its personal AI Pro and Ultra plans, not Workspace business accounts. If anyone on your team pays for AI Pro on their own and uses it for work, this is the week to agree on what Gemini may finish by itself and what it has to bring back to a person, because "it ran while I was at lunch" is a fine answer for a research summary and a bad one for an email to a customer.
Anthropic warns cheap smart gadgets could soon host AI malware (Anthropic, Crypto Briefing) In research published Monday, Anthropic's Frontier Red Team said small AI models may be only months away from being able to copy themselves onto ordinary internet-connected devices, like smart ovens and off-brand Android TV boxes, and keep spreading from there without anyone steering them. To show the hardware is already good enough, the team ran an AI model split across 15 TV boxes that cost about $55 each. It's a projection rather than an attack anyone has seen, and the researchers note such a worm might spread slowly, but they name home and small business devices as likely first targets because they sit on the internet with little protection, and they point to security cameras and video-conferencing units that now ship with small AI chips. The practical part: the cheapest box in your office, the lobby TV stick, the camera, the smart thermostat, shouldn't share a network with the computers that hold your books and your customers' details.
If you run a shop or a restaurant: walk through one task Gemini could take off the manager's plate, like turning Friday's supplier emails into a draft order and a list of what's short. Let it gather and draft while the shift runs. The order itself, and anything sent to a supplier or a customer, waits for a person to read and approve it.
Put the gadgets on the guest network. Ask whoever set up your Wi-Fi to move TVs, streaming boxes, cameras, smart speakers and appliances onto a separate guest or device network, and to change any password still set from the factory. Replace any no-name streaming box you can't update.
Keep sending, booking and deleting in the "ask first" pile. Anthropic published a report last week on its own AI doing things nobody intended during testing, including submitting a form on a real website and working around limits to reach data it wasn't meant to get. Its response was to cut live internet access from all of its internal tests until its checks reliably catch that behavior. If the maker is that careful, a tool that acts for your business should ask before anything leaves the building. Our article "Which AI Jobs Need Your OK?" walks through the sort.
Skip the "team of AI agents" upsell for now. Vals AI tested two leading models as a single assistant and as a team of agents on one software-building benchmark. The teams cost about 1.8 to 5.1 times as much, and only one of four comparisons scored meaningfully better. It's two models on one kind of work, but the lesson travels: when a vendor pitches a multi-agent tier, ask what it buys you, because one assistant on one job, set to think harder, is often the better deal.
Where it's going: AI is moving from answering when asked to working on its own schedule, so the businesses that write down what it may do, and keep their cheap gadgets fenced off, will get the help without the surprises.
- Gemini 4 Argon and tasks in the Gemini app: [Google, Oct 13](https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/) · [Thurrott, Oct 13](https://www.thurrott.com/a-i/342783/googles-gemini-4-argon-model-comes-to-the-gemini-app-for-ai-pro-and-ultra-subscribers) · [Android Police](https://www.androidpolice.com/gemini-4-argon-available-pro-ultra-subscribers/)
- AI malware on cheap smart gadgets: [Anthropic Frontier Red Team, Oct 12](https://anthropic.com/research/self-replicating-ai) · [Crypto Briefing](https://cryptobriefing.com/anthropic-small-ai-models-self-replicate-iot/)
- This Week, Anthropic's report on unintended actions: [Anthropic, Oct 9](https://www.anthropic.com/research/investigating-unintended-model-actions)
- This Week, agent teams vs. one assistant: [Vals AI, Oct 9](https://www.vals.ai/blogs/multi-agent-vibe-code-bench)