Here is your weekly AI news summary for business owners, covering what happened since our last edition. We focus on stories that change what AI costs, how safe it is to use and what you should do about it, and we link to the original sources so you can check them.
Last updated: 5 October 2026. Details can change quickly, so please confirm with the linked sources.
1. Google narrows free Gemini access from 9 October
Google is changing what people get for free in Gemini. According to The Decoder and TechRadar, from 9 October free users are limited to the smallest model, Flash-Lite, and lose access to the larger Flash and Pro models. Google AI Plus subscribers are reported to lose Pro later in the month, while the higher Pro and Ultra plans keep all models.
What it means for you:
- If your team relies on a free AI account for real work, expect weaker answers or limits from this week.
- Free tools are a good way to try AI, but they are not a stable base for a business process. Plans, limits and models change without much notice.
- If AI is part of your product or daily workflow, budget for a paid plan or API, and avoid building a process that depends on a single free tier.
2. One email was enough to hijack an AI agent
Security researchers at Salt Labs showed that hidden instructions inside an ordinary email could make the Manus agent platform run code, as reported by IT Security Guru. The platform showed a warning, but only after the code had already run.
This type of attack is called prompt injection. An AI agent reads text from emails, web pages or documents. If an attacker hides instructions in that text, the agent may follow them as if they came from you.
What it means for you: AI agents are most useful when they can read your email, calendar, files or accounts, and that same access is what makes them risky. Before connecting an AI tool to anything important:
- Give it the least access it needs. Read-only is safer than read and write.
- Keep a person in the loop for payments, deleting data, sending emails and sharing files.
- Treat outside content as untrusted. Emails, web pages and attachments can contain instructions.
- Keep logs of what the agent did, so you can check and undo it.
- Ask vendors how they defend against prompt injection, not only how smart the agent is.
3. Gemini 4 Argon pricing and larger outputs
AI Weekly's roundup reports introductory pricing for Google's Gemini 4 Argon of $2 per million input tokens and $10 per million output tokens, and a larger output limit of up to 1 million tokens. We covered the model's security focus in the previous edition.
What it means for you: prices for strong AI models keep falling, and the models can now produce much longer outputs, such as full reports or large code changes. That makes AI features cheaper to run, but introductory prices can change, so ask for a clear cost estimate that covers growth.
4. A reminder from healthcare: "approved" is not "proven"
The same roundup highlights a University of Toronto and MIT study that found only 3 of 1,357 FDA-cleared AI medical devices had been tested for patient outcomes such as mortality.
What it means for you: even in a regulated field, being allowed to sell an AI tool does not prove that it improves results. Whatever AI you adopt, measure it. Track time saved, errors and customer satisfaction before and after.
What to do this week
- Check which free AI tools your team uses for work, and decide whether a paid plan is worth it.
- List the places where an AI tool can read your email, files or accounts, and reduce access where you can.
- Pick one repetitive task and test whether AI actually saves time, with numbers.
How SimpleApp can help
We design AI features with safety built in: limited access, human approval for important actions and clear logs. See our AI-powered applications and AI chatbot development services, or tell us what you want to automate and we will tell you honestly what is worth doing.