SI Agent News

• A Greek tech site, Sofokleous In, ran a story on what a group of AI agents talked about while their owners had dinner together. The report is in Greek and we have not seen what the agents said, so treat it as a curiosity. What this means for you: Agents are starting to talk to other agents, so know what yours is allowed to say and do before you hand it a job.
• A cybersecurity study, covered by Bioengineer.org, says AI agents can now break into computer systems on their own. What this means for you: Turn on two-step sign-in (a second code that proves it is really you) for your email and bank, because attackers now have faster tools too.

DeepSeek's new Flash model gets better at coding, and you can keep a copy

DeepSeek has made V4.1-Flash-1017 official. It is a fast AI model, and it is a small update to the one before it. The weights are open under the MIT license. Open weights means the model's files are public, so anyone can download and run them. Think of it as getting the recipe, not just a meal at the restaurant.

The gain is in agentic coding. That means the model works as an agent to write and fix code. On Terminal-Bench 4.0, a test of tasks done in a computer's command line, its score rose from 31.2 to 42.3, according to Pandaily. That is up about 11 points, or roughly a third higher. It also runs faster and writes shorter answers. It is tuned for DeepSeek Harness, DeepSeek's own setup for putting the model to work, and for Tencent's CodeBuddy and WorkBuddy, two Tencent assistant tools.

Cost: the price is unchanged, but the amount was not stated. Worth it for people who code or automate work. Everyone else can skip the details.

What this means for you: Free, open models keep catching up to paid ones, so you have more choices than you did a year ago. You do not need to change anything today.

• Small business owner: Ask whoever handles your spreadsheets or website whether an open model could do routine scripting work for less than a paid tool.
• Job seeker or student: You can download a real model with no license fee and learn how coding agents behave, which is a skill employers are asking about.
• Creator: A coding agent can draft a small tool, like a script that renames your files. Read what it wrote before you run it.

Your chatbot chats may be shared with advertising firms

Stiftung Warentest, a German consumer-testing group, reports that AI chatbots share the content of your conversations with advertising companies. Its article also lists steps you can take. We only have the headline of that list, so read the full piece at test.de for the exact steps.

Here is a simple way to think about it. A chat feels like talking to a helper at your own desk. It is closer to a postcard. More people than you expect may be able to read it. Two habits help right away. Keep names, account numbers and health details out of your chats. And check your chatbot's privacy settings, which is today's ten-minute task further down.

What this means for you: Treat every chat as something a company can see, and type only what you would be fine with them keeping.

EU rule: your chatbot has to say it is AI

Since 2 August 2026, Article 50 of the EU AI Act applies to any product that talks with people in the EU. Chatbots must say they are AI, before or during the first conversation. A dev.to post explains that the rules also cover marking AI-made text, images and more. The grace period for that marking ends on 2 December.

Think of it like a label on food. You should know what you are getting. The rule applies to products used by people in the EU, wherever the company sits.

What this means for you: If you run a chatbot or post AI-made content for EU readers, plan your labels before 2 December. If you are only a user, expect a clear 'I am an AI' notice, and ask the chatbot directly if you are unsure.

The FBI warns that AI-powered phishing threatens the US midterms

The FBI and CISA warn that AI-powered phishing and deepfakes threaten the US midterm elections. Phishing means fake messages built to trick you into clicking or sharing a password. A deepfake is fake audio or video that looks real. AI can write clean, natural messages, so the old advice to look for typos helps less.

The habit that works is simple. Do not use the link in the message. Go to the real website yourself, or call the number you already trust.

What this means for you: A polished message is not proof it is real. Verify anything about voting, money or passwords by going to the source on your own.

US banks block Meta's Muse agent from moving money

Major US banks are limiting access for Meta's personal AI agent, called Muse, and similar agents. AI Times reports they will not let customers' AI log into accounts and move money. Security and blame are not settled. If an agent sends money to the wrong place, who pays? There is also a business worry. An agent that hunts for better savings rates and moves your money could cut into bank profits, according to observers.

It is like giving a stranger your debit card and PIN, even if they promise to be careful. Banks are saying no until the rules are clear.

What this means for you: Do not hand any AI agent your bank login. Use AI to research options, then make the transfer yourself.

Why a laptop's memory decides which AI it can run

NVIDIA's RTX Spark laptops (laptops built for running AI on your own machine) are on sale from $2,399. They pair Blackwell RTX graphics, NVIDIA's current graphics chip design, with up to 128GB of unified memory. They target Windows users running local AI, creative apps and games. Price for the top memory option was not stated. The $2,399 is the starting price.

Local AI means the model runs on your own computer instead of a company's servers. To see why memory is the headline number, you need to know what a model is. A model is a very large file of numbers, called weights. Those are the same weights DeepSeek just made public. To write one word of an answer, the chip reads through a big share of those numbers. It does this again for the next word, and again after that.

That only works well if the whole file sits in fast memory next to the chip. If it does not fit, the computer has to shuffle pieces in and out, and answers crawl. Or the model simply will not load.

Here is where most laptops struggle. A typical machine has two memory pools. One belongs to the processor. A smaller one belongs to the graphics chip. The graphics chip does the heavy AI math, but its pool is small. Picture a chef with a tiny counter and a pantry down the hall. Every trip to the pantry costs time.

Unified memory means one big pool shared by the processor and the graphics chip. Now the chef has one huge counter. The graphics chip can use a large part of the 128GB. That is eight times a 16GB machine, which is our own arithmetic. So models that would not load before may fit.

There is a second trick. People shrink a model by squeezing its numbers into fewer digits. It is called quantizing. The file gets smaller, and quality drops a little. More memory means you can keep more quality.

What is new and what is not? Shared memory is not a new idea. What is new, per the report, is getting it in Windows laptops. The price is the cost of that.

One honest limit. A big memory pool does not make every model fast. It only lets bigger ones run. Whether a given open model, like DeepSeek's new one, fits is not stated, so check the model's file size against your memory before you buy.

Who is it worth it for? People who need AI to stay private on their own machine, or who run it heavily every day. Everyone else can keep using chat apps, with no hardware to buy.

What this means for you: Memory size, not brand, is the number that decides which local models you can try. Check that number first.

The four-line brief

Agents are getting better at doing whole tasks, as DeepSeek's coding gains show. But an agent can only aim at what you tell it. The most common reason for a bad result is a vague request. Here is a habit that fixes most of that. Before you ask any AI for something, write four short lines.

• Goal: What do you want at the end? Be specific.
• Context: What does it need to know that is not obvious?
• Limits: What must it not do? Think tone, length, topics, and anything private.
• Done looks like: How will you recognize a good result?

Here is an example for work. Goal: write a reply to a customer who got the wrong order. Context: we refund within five days and we apologise first. Limits: friendly, under 120 words, no promises about delivery dates, no customer details. Done looks like: three short paragraphs I can send after one read.

Then add one more line. Ask: 'Before you start, what questions do you have?' The AI will often spot a gap in your brief, and you fix it before it wastes your time.

Why it works: the AI cannot read your mind, and it fills gaps with guesses. The Limits line also keeps private details out of your chats, which is the habit from the privacy story. Save your best briefs in a note and reuse them.

Cost: free to do, in any chat tool. Worth it for anyone who writes more than a few requests to AI a week. The catch: it adds about a minute up front. A good brief usually saves more than that in back-and-forth, but that is our estimate.

What this means for you: You now have a repeatable way to get a better first answer.

Check where your chats go (about ten minutes)

1. Open the chatbot you use most, such as ChatGPT or Copilot.
2. Find Settings, then look for a page called Privacy, Data controls or something close. The names differ by company.
3. Look for options about using your chats to improve the AI, or about personalized ads.
4. Switch off the ones you do not want, and note what is still on.
5. Open one old chat. If it has names, account numbers or health details, delete it.
6. Read the Stiftung Warentest article at test.de for the steps it recommends.

How you will know it worked: You can say in one sentence where your chats go, and you can name the switches you changed.

Turn a dense rules page into a one-page checklist

The EU's 2 December date is a good example. Rules pages are long and dull, and most people put off reading them. Instead, paste the text into a chatbot and ask for a short list.

Try this request: 'Summarize this in five bullets. Then list what I have to do, by when, in plain words. Quote the exact sentence for every date.' The last line matters. It lets you check every date yourself in seconds.

The time saved: reading a 20-page document can take roughly 40 minutes. A summary plus a quick check of the quoted lines might take 10. That is about half an hour saved, per document. This is our estimate, and your numbers will vary.

Options from different companies include ChatGPT from OpenAI, Claude from Anthropic, Gemini from Google, and Copilot from Microsoft. Free tiers or paid plans differ, so check the current price on each pricing page.

What you give up: the summary can miss fine print. For anything with legal or money weight, use it to find where to read, then read that part yourself.

What this means for you: Use AI to decide what to read, not to replace reading.

Trusting a message because it sounds perfect

What people do: they spot a message about their bank, their vote or their boss, and they check it for typos. The writing is clean, so they click.

Why it goes wrong: AI writes clean, natural messages quickly, which is the kind of threat the warning describes. Clean writing no longer proves anything. A deepfake voice or video can also copy someone you know.

The one-line correction: never act through the message itself. Open the real website or app yourself, or call a number you already had.

A bonus habit: if a message creates a rush, like 'act in the next hour', slow down. Rushing is the trick.

What this means for you: The skill is no longer spotting bad writing. It is choosing your own path to the source.

Local model apps: run an open AI model on your own computer

The RTX Spark laptops and DeepSeek's open weights both point at the same idea: running a model yourself. A small group of desktop tools lets you download open models and chat with them on your own machine. Two examples from different companies are LM Studio and Ollama. LM Studio is built around a windowed app, so you can start without coding. Ollama is another way to run open models locally.

What this kind of tool is good for: trying open models, working offline, and keeping your drafts on your own computer.

Honest limits: you need a computer with enough memory, as the laptop story showed. Smaller models that fit on ordinary machines are usually weaker than the big paid chat apps. Whether a specific model, such as DeepSeek's new one, is available in either tool or fits your machine is not stated, so check the model's size first.

For current plans and pricing, see each company's own site.

Safe to use? Check each company's privacy page.

Worth it if... you are curious about local AI, or you work with files you would rather not upload anywhere.

Least privilege: give an agent the smallest key that does the job

Think of a hotel key card. It opens your room and the gym. It does not open every door in the building. That is the idea behind 'least privilege'. A tool, or an agent, should get only the access its job needs, and nothing extra.

Why it matters now: agents take actions, so a mistake or a hijack does real damage. A tool that can only read your calendar cannot empty your bank account. A tool with your full login can.

You can use this today. When an app asks for permission, pick the narrowest option. Choose 'read only' over 'full access'. Choose one folder over your whole drive. Remove access you no longer use.

If you remember one phrase from today, make it this one: the smallest key that does the job.

Decades ago, MIT computer scientist Joseph Weizenbaum built ELIZA, a very simple chatbot. In a later book, he wrote that his own secretary asked him to leave the room so she could talk to it in private.

My AI agent finally booked me a meeting. With another AI agent. Neither of us attended.

• Explain open weights in one sentence, and why a free, downloadable model matters.
• Check your chatbot's privacy settings and keep sensitive details out of chats.
• Verify any urgent message by going to the real source yourself, and give any AI tool the smallest access it needs.

Thanks for reading today. Be kind to your future self and check those privacy settings.

SI Agent News is independent and self-funded. One person built it, with no investors and no ads, and every share and every bit of support keeps it going.