Products SEO & GEO Agent Paid Ads Agent Website Agent Conversion Agent
How it works Pricing FAQ Blog Book a call

AI Agent Safety: How to Limit What Your AI Agents Can Touch

Updated · 8 min read

Short answer Give each AI agent its own login with the lowest access level that does its one job, cap what it can spend or send, and make it ask you before anything you can't undo: payments, deletions, new campaigns, mass emails, adding users. Don't let one agent read outside content, see private data and act on its own at once.

Want help applying this to your business? Talk it through on a free 30-minute call.

Book a call

How do you limit what an AI agent can touch?

Set three limits for every agent: what it can reach, what it can change, and when it has to stop and ask. Security people call the first one "least privilege": the smallest access that still gets the job done.

LimitWhat it meansHow to set it
ReachOnly the accounts and data its job needsIts own login, added as a user; no shared owner password
ChangeThe lowest role that does the workRead-only first; edit rights only where it must act
AskA person approves anything costly or permanentApproval step in the tool, plus spend and send caps

The OWASP GenAI Security Project, a nonprofit security community, names the problem "excessive agency" and traces it to three causes: too many functions, too many permissions and too much autonomy (acting without a check). Its fix list starts here:

"Limit the extensions that LLM agents are allowed to call to only the minimum necessary." (OWASP: LLM06:2025 Excessive Agency)

An LLM (large language model) is the engine behind ChatGPT, Claude and Gemini. An "extension" or "tool" is a connection that lets it read from or change another system, such as your ad account or inbox. If you're new to agents, start with agentic AI explained.

Why give an agent less access than you'd give a new hire?

Because an agent can be fooled by text a person would ignore, and it repeats a mistake as fast as it works. The main trick is prompt injection: instructions hidden inside an email, review or web page the agent reads. Anthropic, which makes Claude, describes them as:

"adversarial instructions hidden within the content that AI models process" (Anthropic, November 2025)

Its example is an email with instructions "embedded in white text, invisible to you but processed by the agent." Even after its defenses cut attacks on its browser agent to about 1% success, Anthropic says that rate "still represents meaningful risk" (same source). No setting removes the risk. Smaller access limits how much damage a fooled agent can do: an agent with no delete button can't delete your files, whatever it reads.

The Rule of Two: a quick test for any agent

Meta published a simple test in October 2025. An agent may have at most two of these three powers in one session (Meta AI: Agents Rule of Two):

  1. It reads content you don't control: customer emails, web pages, reviews, uploaded files.
  2. It can see private data or sensitive systems: customer lists, payment details, admin settings.
  3. It can change things or send messages out: publish, spend, email, delete.

Some summaries describe the third power as the agent "learning on its own." Meta's text doesn't say that; it's about changing state or communicating outward. When a job truly needs all three, Meta says the agent "should not be permitted to operate autonomously" and needs human approval or another reliable check.

Agent jobOutside contentPrivate dataActsVerdict
Drafts replies to Google reviews for you to postYesNoNoFine alone
Reads your ad reports and suggests changesNoYesNoFine alone
Answers web chat and books into your calendarYesYesYesNeeds approvals or narrow limits
Reads your inbox and sends replies itselfYesYesYesNeeds approvals

Which access level to give, account by account

Add the agent as its own user wherever the platform allows it. Then you can see what it did, change its role, and remove it without resetting your password.

AccountRole to start withWhy
Google AdsRead-onlyCan "view campaigns" and run reports, but not edit (Google Ads Help)
Google Ads, when it must actStandardCan "edit campaigns", but also "edit billing information"; it can't add users (same source)
Google AdsNever AdminAdmin can "give account access, change access levels" (same source)
Google Business ProfileManager, never OwnerManagers can't add or remove users or delete the profile (Business Profile Help)
EmailDrafts only, or a separate sending addressA bad send to your whole list can't be recalled
WebsiteA page-editing role, not administrator, if your site builder has oneKeeps settings, users and site deletion out of its reach
Bank, payment and payrollNo accessNo marketing job needs it

Note the Google Ads catch: Standard access includes billing edits. If that worries you, keep the agent on Read-only and have it hand you a list of changes to apply.

Cap the money too. In Google Ads, a campaign can spend up to two times its average daily budget on a given day, but the monthly charge is capped at "30.4 times your average daily budget for most campaigns" (Google Ads Help: About average daily budgets). So the budget you set is the agent's spending ceiling. Keep it at a figure you'd accept losing in a bad month. More on letting AI move ad money is in AI agents for ad spend.

What should an AI agent never do without you?

Anything you can't undo, anything that moves money, and anything that changes who has access. OpenAI's guide for businesses building agents puts it this way:

"Actions that are sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent's reliability grows." (OpenAI: A practical guide to building agents)

Its examples are canceling orders, large refunds and payments. For a small business's marketing, the list looks like this:

  • Pay, refund, change billing or raise a budget.
  • Launch a new campaign or turn on a new ad platform.
  • Delete pages, campaigns, contacts or files.
  • Email or text more than a handful of people at once.
  • Add, remove or change users and passwords.
  • Publish anything that states a price, a guarantee, a medical, legal or financial claim, or a promise to a customer.
  • Reply to an angry customer or a public complaint.

Everything else can run, as long as you can see a record of it afterward.

How to set approvals without slowing everything down

Rate each action by risk once, then let the low ones run. OpenAI suggests rating each tool low, medium or high "based on factors like read-only vs. write access, reversibility, required account permissions, and financial impact" (same guide).

RiskExamplesRule
LowRead reports, draft posts, suggest keywordsRuns alone
MediumPause a keyword, post a draft-approved update, reply to a 5-star reviewRuns, you check the log weekly
HighSpend, delete, mass send, publish claims, change usersWaits for your yes

Two more brakes: rate limits, a cap on how many actions per hour (from the OWASP list), and failure limits. OpenAI says to cap retries and hand the job to a person when the agent keeps failing. Then read the activity log every week, the same way you'd skim a new employee's sent folder. To decide which jobs to automate at all, see AI marketing automation: what to automate and what humans own.

Who is responsible when an agent gets it wrong?

You are. The agent acts in your name, and the rules for ads, emails and claims still apply to what it publishes. The US Federal Trade Commission said in its 2024 crackdown on deceptive AI schemes:

"The FTC's enforcement actions make clear that there is no AI exemption from the laws on the books." (FTC, September 2024)

That's one more reason customer-facing claims sit on the never-alone list.

Example

A plumber's ad agent

The agent gets Read-only in Google Ads and a $50 average daily budget, so the month can't pass about $1,520 (50 × 30.4). Each Monday it sends a list of wasted search terms; the owner approves them in five minutes and applies them. Nothing it reads can make it spend more.

A dental office's review replies

The agent is a Manager on the Business Profile, not an Owner, so it can't remove users or delete the listing. It posts thank-you replies to 4- and 5-star reviews alone. Anything below 4 stars goes to the office manager as a draft, because a reply about a patient's care can disclose health details.

A landscaper's email agent

The agent writes the monthly newsletter and loads it as a draft for 2,000 contacts. The owner reads it and presses send. The agent never holds the sending password.

FAQ

Can I just give the agent my own login?

Better not. You can't tell its actions from yours or lower its access alone, and a breach of the tool exposes your full-access password.

Is Read-only access enough to be useful?

Often, yes. It can find waste, draft changes and write reports for you to apply. Give edit rights after weeks of correct suggestions.

Does a good prompt keep an agent safe?

It helps, but text the agent reads can override it. The platform enforces permissions and caps, so they hold even when the agent is fooled.

How often should I review what my agent did?

Weekly for a new agent, monthly once it has a clean record. Remove its access the day you stop using the tool.

Different for you?

Keep learning

NextAI assistant data: what to share and what to keep privateThe smallest set of data each AI task needs, the lowest account role to grant, and what never goes into ChatGPT, Gemini or Claude.

Ask us one question

One line is enough. We reply by email.

Sources

  1. OWASP GenAI Security Project: LLM06:2025 Excessive Agency
  2. Meta AI: Agents Rule of Two (October 2025)
  3. OpenAI: A practical guide to building agents
  4. Anthropic: Mitigating the risk of prompt injections in browser use (November 2025)
  5. Google Ads Help: About access levels in your Google Ads account
  6. Google Ads Help: About average daily budgets
  7. Google Business Profile Help: Manage your Business Profile owners & managers
  8. FTC: FTC Announces Crackdown on Deceptive AI Claims and Schemes (2024)