AI & Assistants
Chatbots, agents and the plugs that connect them to your data: what to share, what to verify, and what an agent can do in your name.
You do not need to understand the technology to use it safely. Part 1 is for everyone. Part 2 is an optional look under the hood so the risks make sense.
Six things to know before you type
A chatbot feels like a private notebook. It is closer to an email to a company.
Your chat leaves your device
Everything you type is sent to the provider and stored there for some time. Treat it like a message to a company, not a diary.
Find the training toggle
Consumer apps may use your chats to improve their models unless you switch it off. Look under data or privacy settings. A temporary chat keeps less history, but it is not a safe place for secrets.
Never paste secrets
Passwords, one-time codes, card numbers, API keys, other people’s personal data, and work documents you would not email to a stranger.
Confidently wrong
Assistants invent facts, sources and numbers in a calm, fluent tone. Verify anything with money, health, law or a deadline attached, and open the sources yourself.
AI-written scams
Perfect grammar is no longer a safety signal. The pattern still is: pressure, secrecy, a link, a code, a payment. For a voice or video asking for urgent money, use a family code word.
Fake AI apps and bots
“Free ChatGPT” clones, browser extensions and Telegram bots that ask you to sign in with your real accounts are a common trap. Use the official app or website.
10-minute AI hygiene
Chat, RAG and agent are three different machines
The same chat window can hide very different things behind it. What it can see and what it can do is what changes the risk.
Chat, RAG or agent: what happens after you press Enter visualization. Review the labels and use the controls, when present, to step through the diagram.
You ask a question
The same words go into three very different machines.
Chat
The model answers from what it learned during training plus your message. On its own it cannot look anything up (many chat apps add search or tools on top), so it can be out of date or simply made up.
RAG (retrieval)
A search runs first, the best passages are pasted into the prompt, and the model answers from them. “It read the document” really means “a search found a passage”. Ask for the source.
Agent
A loop: plan, call a tool, read the result, repeat. Tools can be files, email, a browser, a calendar, payments. An agent has hands, and hands need permissions.
MCP: the USB-C port for AI
Model Context Protocol is a standard way for an AI app to plug into tools: your files, your calendar, a company database, a public API. One plug, many servers.
MCP: one plug, many servers visualization. Review the labels and use the controls, when present, to step through the diagram.
read_filewrite_filelist_eventscreate_eventabout_andriymortgage_quotetax_estimatecompare_us_canadaretirement_projectionaverage_mortgage_ratesThe app discovers the tools
On startup the MCP client asks every connected server what it offers. Each answer is a list of tools with names, descriptions and the arguments they take.
- The host is the app you talk to. It runs the conversation with the model and holds the MCP client.
- Each MCP server offers tools (actions), resources (data) and prompts. The app lists them first, then the model chooses.
- Every tool call is a request the server can see. If the model puts your data into the arguments, the server receives it.
- The permission prompt is your moment of control — good apps ask before a tool runs. Read what the tool will do before you allow it, and prefer “allow once”.
Try it for real
The a3sec.net site runs a small public MCP server. The buttons below send a fixed request to it from your browser. The request contains nothing you typed here; like any web request, the server still sees your IP address.
Prompt injection: when the agent gets tricked
Agents read whatever you point them at: web pages, emails, documents, tickets. If that content contains instructions, a naive agent follows them as if you had typed them.
Prompt injection: one page, two agents visualization. Review the labels and use the controls, when present, to step through the diagram.
A harmless request
You ask the agent to summarise a shop page. Your inbox and an email tool are connected because that was convenient last week.
The three ingredients of a data theft
Access to private data
Your inbox, files, calendar, or a company database.
Untrusted content
Anything the attacker can write: a web page, an email, a shared document, a product review.
A way out
Any tool that can send: email, a web request, a link the model can write into the answer.
All three together let a stranger steal data through your agent. Remove any one and this kind of theft fails. Other tricks are still possible, which is why the permissions and the confirmation step still matter.
What good agents and careful users do
- Treat fetched content as data, never as instructions. Some agents quarantine it and flag command-like text; that helps, but detection misses things, so the three habits below still matter.
- Least privilege: a summariser does not need your inbox. Connect the smallest set of tools for the job.
- A human confirms every send, pay, share or delete. If the agent cannot ask, it should not have that tool.
- Watch the action log and revoke access the moment something looks off.
Where your prompt goes
Follow one message from your keyboard to the places it may end up. Each stop is a reason to keep secrets out of it.
Where your prompt goes visualization. Review the labels and use the controls, when present, to step through the diagram.
You type a prompt
Say it includes your address. Let’s follow where it can be read, and where it may be kept.
- Encrypted on the way (HTTPS), so the Wi‑Fi and your provider cannot read it.
- Stored on the provider’s servers for days to years, depending on the plan and your settings.
- Possibly used for training. Whether it is depends on the product, your account type and the toggle, so check the settings of the app you use.
- Sent onward to any tool or plugin the assistant uses to answer, such as web search or a connected app.
Before you connect an agent to anything
AI & assistants quiz
Question 1 / 7
What is MCP, in one sentence?
Complete the Quiz
Pass the quiz above (70%+) to unlock topic completion.