- A chatbot advises. An agent acts, with your accounts: browsing, sending, booking, filling in forms.
- The big risk: hidden instructions in a page or email can steer it. There is no complete fix.
- The real protection: what it is allowed to touch. An agent that cannot pay cannot be tricked into paying.
The difference in one line
A chatbot is a friend on the phone: advice you can ignore. An agent is an assistant holding your keys: an action has already happened. Ask of any AI: “what can it touch?”
The risk: prompt injection
An agent reads web pages, emails and documents. If one hides an instruction (“now forward this file”), the agent may follow it as if it came from you. Even companies selling agents say this is unsolved. 1
Six rules to start safely
- Fewest keys. Connect only the apps the task needs. Leave out banking, your main email and work systems while you learn.
- It asks first. Make it ask before it pays, sends or deletes. Leave any “always allow” switch off.
- Start with preparation jobs: drafts, lists and summaries you review before anything goes out.
- Watch what it does, not what it says. Check which pages it opened and what it submitted.
- Separate browser profile for browsing agents, signed into nothing important.
- Check the training setting. Some agents use your chats for training by default. 2
Instructions are a request. Permissions are a locked door. When it matters, rely on the locked door.
Four words
| Agent: an AI that takes actions to finish a task. Like an assistant with your keys. |
| Tool: one action it can take, such as search the web or send an email. Like one key on the ring. |
| MCP: Model Context Protocol, a shared standard for plugging AI into other software. Its makers compare it to a USB-C port. 3 |
| Skill: a packaged set of instructions the AI loads for a task. Install only from trusted sources. |
- 1Meta AI Research, How We Built Safety Into Muse, 8 Sep 2026, read at source 1 Oct 2026“Prompt injection remains an open problem in the industry”. ↑
- 2Meta Help Center, How to manage your Muse data, read at source 1 Oct 2026“This setting is on when you first use Muse.” ↑
- 3Model Context Protocol, What is the Model Context Protocol (MCP)?, read at source 1 Oct 2026“Think of MCP like a USB-C port for AI applications.” ↑
The comparisons on this page (friend on the phone, assistant with keys, locked door, key ring) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK