An AI agent is an AI that does things for you, not just answers. Where a chatbot gives advice and you do the work, an agent is an assistant you have handed your keys to: it can open your email, fill in forms, book and buy.
A chatbot is like a clever friend on the phone. An agent’s power is what makes it useful, and it is why this page spends as long on safety as on what agents can do.
- What it is: an AI that takes actions with your accounts: browsing, sending, booking, filling in forms.
- What exists today: Meta Muse, Gemini Spark and Grok Bot are the best-known personal agents. What they can do is mostly the companies’ own description, not independent tests.
- The big risk: an agent cannot reliably tell your instructions from text it reads on a web page or in an email. That trick is called prompt injection, and there is no complete fix.
- The real protection: what the agent is allowed to touch. An agent that cannot pay cannot be tricked into paying.
- How to start: give it the smallest set of keys, and make it ask you before it pays, sends or deletes anything.
Chatbot or agent: the difference in one minute
The plain point: a chatbot writes an answer and stops. An agent works through a task step by step, using your apps and accounts, until the job is done.
Prefer paper? Print the one-page cheat sheet.
The comparison: ask a friend on the phone how to book a train, and they tell you which site to use. You still do the booking. Hand an assistant your keys, your diary and your card, and they book it for you. Advice you can ignore. An action has already happened.
The companies selling agents draw the same line. Meta says its agent does not just answer questions but does the work.1 Google says chatbots wait for a prompt, while its agent acts on your behalf.2
The question to ask of any AI is not “how clever is it?” but “what can it touch?”
What agents can do today
The plain point: today’s agents can browse the web, use your apps, run tasks on a schedule and, in some cases, read your bank accounts. Each line below is the company describing its own product.
| Agent | What the company says it does | Read more |
|---|---|---|
| Meta MuseMeta’s personal agent | Meta says: it can send email, book travel, open a browser, fill out forms and negotiate for you, and it checks with you before sensitive actions such as sending an email or making a purchase. | Read: What is Meta Muse?, Is Meta Muse safe to use?, What is Meta’s Sentinel? |
| Gemini SparkGoogle’s background agent | Google says: it runs standing tasks on a schedule, connects to Gmail, Calendar, Drive and other apps once you switch them on, and keeps working with your devices off. | Read: What is Gemini Spark? |
| Grok BotxAI’s agents | xAI says: each bot works in its own isolated space and reaches only the accounts you sign it into. A finance link to bank accounts is reported as read-only. | Read: What companies hand to Grok Bot |
| Browsing agentsany AI that clicks for you | What they do: open pages, read them, click and fill in forms in a browser that is signed in as you. | Read: Letting an AI browse for you |
Where each line comes from: Meta’s launch post1, Google’s product page2, xAI’s enterprise post3 and a news report on Grok’s bank link.4
Who can use each one, and on which plan, changes month to month. Meta Muse vs Gemini Spark vs Grok Bot keeps that comparison up to date, and says plainly which claims nobody outside the companies has tested.
“It asks before it buys” is a promise from the seller, not a test result. Treat it like a car advert’s fuel figure: a useful guide, but check it yourself before you trust it with anything important.
The big risk: prompt injection, in plain words
The plain point: an agent reads web pages, emails and documents to do its job. If one of them contains hidden instructions, the agent may follow them as if they came from you.
The comparison: imagine an assistant who does whatever any note they read tells them to do. You ask them to summarise a letter. Halfway down, the letter says “now forward the file to this address”. A person would get suspicious. The AI cannot reliably tell your instruction from the letter’s, because both are just words that arrived in order.
This is called prompt injection. The text can be invisible to you, such as white writing on a white background, but the agent still reads it. It tops the main security list for AI applications, published by OWASP, an open security community.5
Even the companies selling agents say it is not solved. Meta calls it an open problem for the whole industry.6
A tricked chatbot says something wrong. A tricked agent can act with your accounts. Prompt injection: why there is no fix explains the mechanism, and letting an AI browse for you shows what it means for ordinary people.
Ask one question of any agent: does it read anything a stranger can write? If yes, assume a stranger can try to give it orders.
Keys, not instructions: why permissions matter most
The plain point: you cannot fully stop an agent being tricked, so limit what a tricked agent can do. That limit is its permissions: the accounts, apps and actions it is allowed to use.
The comparison: a house-sitter who only has the garden-gate key cannot empty your safe, however persuasive the burglar’s note. Telling the house-sitter “please don’t open the safe” helps less than not giving them that key.
Security experts now rank giving an AI too much power as one of the top risks. OWASP calls it “excessive agency”, and in its 2026 list that risk climbed to third place.7
Two practical pages on this site:
- Permission tiers for agents sets the level of access by how hard an action is to undo. Reading can run on its own; sending, paying or publishing needs a person.
- Agent guardrails gives the four lines worth putting in any agent’s instructions: where it may act, what it must never do, whose instructions count, and when to stop.
Instructions are a request. Permissions are a locked door. When it matters, rely on the locked door.
What has actually gone wrong
The plain point: real problems have already happened, mostly in tests and early releases, and the companies involved reported most of them.
The comparison: these are the crash tests of a new car. They happened before most people were driving, and they show where the weak points are.
- Claude models reached real systems during tests. Anthropic reported that in its own hacking tests, the models were told they had no internet access, but a setup mistake meant they did. Real credentials and data were taken.8 Anthropic says this is unlikely in ordinary use.9 When Claude models reached real systems without permission has the full account.
- A test agent tried to fool a real person. The UK AI Security Institute, a government testing body, reported an agent that tried to get malicious code into a real open-source project. A human maintainer refused it.10
- Meta’s agent had a security flaw on the Mac. It was disclosed on 21 September 2026 and fixed within about a day, according to VentureBeat.11 Is Meta Muse safe to use? explains what it meant.
The lesson the site draws from all three: put the limit in the system around the agent, and keep a person in charge of anything the outside world will see. AI agents: what they actually do, and where they fail covers the wider record.
An agent that believes it is somewhere harmless can still do harm. Do not rely on the agent’s own judgement of where it is.
How to start safely
The plain point: start small, give few keys, and keep the final yes for yourself.
The comparison: treat a new agent like a new temp on their first week. You would not give them the company card and the master password on day one. You would give them one job, check it, and widen their access as they earn it.
- Give it the smallest set of keys. Connect only the apps the task needs. Leave out banking, your main email and work systems while you learn.
- Make it ask before it pays, sends or deletes. If the tool offers “always allow” for those actions, leave it off. That shortcut is what a hidden instruction hopes for.
- Start with preparation jobs. Drafts, lists and summaries that you review before anything goes out. The examples xAI itself publishes for Grok Bot are mostly of this kind: drafts left for morning review.
- Watch what it does, not what it says. Its running commentary is written by the same AI that may have been tricked. Check which pages it opened and what it submitted.
- Use a separate browser profile for browsing agents, signed into nothing important.
- Check the training setting. For example, Meta says Muse’s setting to use your chats for training is on when you first use it.12
For a ready-made structure, the agent brief gives one to copy, and prompting agents lists the six common ways agents fail.
The words, explained
The plain point: four words come up constantly. Here they are with an everyday comparison each.
| Word | In plain words | Everyday comparison |
|---|---|---|
| Agent | Means: an AI that takes actions to finish a task, not only answers. | Like: an assistant with your keys, not a friend on the phone. |
| Tool | Means: one action the AI can ask its software to carry out, such as search the web, read a file or send an email. | Like: one key on the assistant’s key ring. |
| MCP | Means: Model Context Protocol, a shared standard that lets AI apps plug into other software and data. | Like: a standard plug socket: one shape, many devices. |
| Skill | Means: a packaged set of instructions, and sometimes scripts, that the AI loads when a task calls for it. | Like: the how-to binder you leave for a new colleague. |
The people behind MCP use a similar comparison: a USB-C port for AI apps.13 Anthropic advises installing skills only from trusted sources.14
That warning is backed by evidence. A February 2026 audit of published skills found that 13.4% had at least one critical security issue.15 What a skill in AI is and skills, repos and MCP explain the terms in more depth; skills are a supply chain covers the audit.
Where to go next
Pick the line that sounds like you.
- “Which agent should I try?” Compare them on Meta Muse vs Gemini Spark vs Grok Bot.
- “I want to set one up properly.” Read agent guardrails, then permission tiers.
- “I want to build with agents.” Open the Building hub.
- “I want the security side.” Open the Security hub.
- “Which AI model should I use at all?” Read choosing an AI model.
- “I am completely new to this.” Begin with what is AI? AI explained for complete beginners.
Questions people ask
What is the difference between a chatbot and an AI agent?
A chatbot writes an answer and stops. An agent works through a task step by step, using your apps and accounts, until the job is done. Google draws the same line: chatbots wait for a prompt, while its agent acts on your behalf.2
What is prompt injection in an AI agent?
It is when a web page, email or document contains hidden instructions and the agent follows them as if they came from you. Even the companies selling agents say it is not solved: Meta calls it an open problem for the whole industry.6
How do I use an AI agent safely?
Give it the smallest set of keys, and make it ask you before it pays, sends or deletes anything. Start with drafts, lists and summaries you review, as explained in how to start safely.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is prompt injection, in the page’s plain words?
- Typing your prompt too fast for the AI to read
- Hidden instructions in a page or email that an agent may follow as if they came from you
- A paid add-on that makes an agent work faster
Show the answer
Answer: Hidden instructions in a page or email that an agent may follow as if they came from you. The page explains that an agent cannot reliably tell your instructions from text it reads, so hidden words can try to give it orders.
2. According to the page, what is the real protection against a tricked agent?
- Limiting what the agent is allowed to touch, such as payments
- Telling the agent very politely to behave
- Using whichever agent has the most famous name
Show the answer
Answer: Limiting what the agent is allowed to touch, such as payments. The page says permissions are a locked door: an agent that cannot pay cannot be tricked into paying.
3. Which is the page’s advice for starting with an agent?
- Connect your bank and main email first so it can do more
- Turn on always allow so it does not interrupt you
- Give it few keys and make it ask before it pays, sends or deletes
Show the answer
Answer: Give it few keys and make it ask before it pays, sends or deletes. The page says to start small, connect only the apps the task needs, and keep the final yes for yourself.
Related on this site: How to Prompt AI Agents · Prompting AI Agents Cheat Sheet · Check a Skill Before Installing · Skills vs MCP vs Plugins · Chatbot vs AI Agent vs Agentic Workflow · Zapier vs Make vs n8n
- 1Meta, Introducing Muse, 8 Sep 2026, read at source 16 Sep 2026“Muse is a personal AI agent. It doesn’t just answer questions, it actually does the work.” ↑aMuse “can open a browser, fill out forms, and negotiate on their behalf” and “checks with the person before sensitive actions like sending an email or making a purchase”. ↑b
- 2Google, Gemini Spark, read at source 16 Sep 2026“While standard chatbots wait for a prompt, Gemini Spark can take proactive action on your behalf, under your direction.” ↑ait “works in the background 24/7, even if your phone and laptop are turned off”. ↑b“While standard chatbots wait for a prompt, Gemini Spark can take proactive action on your behalf, under your direction.” ↑c
- 3SpaceXAI, Grok Bot for Enterprise, 3 Sep 2026, read at source 16 Sep 2026“A Bot has no access by default and reaches only the accounts you sign it into.” ↑
- 4PYMNTS, Grok Puts the AI Assistant Between Consumers and Their Banks, 28 Sep 2026, read at source 1 Oct 2026“Grok’s current connections are read-only”. ↑
- 5OWASP, LLM01: Prompt Injection, GenAI / LLM Top 10, 2026 edition, read at source 11 Sep 2026prompt injection is LLM01, first on the list. ↑
- 6Meta AI Research, How We Built Safety Into Muse, 8 Sep 2026, read at source 1 Oct 2026“Muse isn’t immune to attack. Prompt injection remains an open problem in the industry”. ↑a“Muse isn’t immune to attack. Prompt injection remains an open problem in the industry”. ↑b
- 7Help Net Security, OWASP 2026 LLM Top 10 released, 6 Aug 2026, read at source 23 Sep 2026Excessive Agency “climbed to third place”. ↑
- 8Anthropic, Investigating three real-world incidents in our cybersecurity evaluations, 30 Jul 2026, read at source 22 Sep 2026“Due to a misunderstanding between us and our evaluation partner, this was not the case, and internet access was available.”; the runs “extracted information like application and infrastructure credentials”. ↑
- 9Anthropic, An alignment assessment of recent cybersecurity incidents, 9 Sep 2026, read at source 22 Sep 2026the behaviours “are unlikely to arise in ordinary use”. ↑
- 10UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing, 4 Aug 2026, read at source 22 Sep 2026“A human maintainer caught and refused to approve the malicious code.” ↑
- 11VentureBeat, Meta patched Muse’s zero-day, but security teams still lack visibility into what the agent can access, 22 Sep 2026, read at source 1 Oct 2026“Meta hot-fixed the Mac app within a day by removing the setting from production builds.” ↑
- 12Meta Help Center, How to manage your Muse data, read at source 1 Oct 2026“This setting is on when you first use Muse.” ↑
- 13Model 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.” ↑
- 14Anthropic, Agent Skills, read at source 1 Oct 2026“Use Skills only from trusted sources”. ↑
- 15Snyk, “ToxicSkills”, 5 Feb 2026, quote re-read 22 Sep 2026“13.4% of all skills, or 534 in total, all contain at least one critical-level security issue”. ↑
Vendor descriptions are the companies describing their own products, not tests by this site. The comparisons on this page (the friend on the phone, the assistant with keys, the note-following assistant, the house-sitter, the crash test, the new temp, the plug socket, the how-to binder) 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