# How do you secure AI agents?

> Limit what each agent can do, decide each action outside the AI model, read what it reads and keep a record you can check. Verosek does all four in one place, so security is not a second vendor.

Field guide · Vats Shah · 6 November 2026 · 7 min read
Canonical: https://verosek.com/blog/how-do-you-secure-ai-agents

## In short

You secure an AI agent by limiting what it can do, because you cannot count on what it decides. Give each agent narrow access, decide each action outside the AI model, read what the agent reads and keep a record you can check. Verosek does this in one place: it sits between your agents and your systems, and your rules decide each tool call before it runs. Shield, its scanning layer, comes with the gateway and reads the text once you turn it on, so there is no second vendor to add.

Secure what an agent can do, not only what it says. A tool call is how an agent reaches one of your systems: it reads a row, writes a record or calls an API. A wrong idea becomes a real change at that call, so start there if your own agents write to production.

## What is AI agent security?

**AI agent security means keeping an agent's actions inside limits you set, and being able to show what it did.** AI security is the wider field: models, their data, the text going in and out. Security for AI began with reading words, because words are all a chatbot produces. An agent also acts. OWASP, the security group, wrote a separate [Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) for that reason, and [its launch post](https://genai.owasp.org/2025/12/09/owasp-top-10-for-agentic-applications-the-benchmark-for-agentic-security-in-the-age-of-autonomous-ai/) says "Once AI began taking actions, the nature of security changed forever".

*Figure 1. The same bad idea, in a chatbot and in an agent. Illustrative values.*

## What are the risks of AI agents?

**Most of them come down to an agent doing more than it should, for someone it should not, with nothing to show afterwards.** OWASP calls the general failure Excessive Agency, and says its root cause is "typically one or more of" three things: "excessive functionality; excessive permissions; excessive autonomy" ([OWASP LLM06](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)). That holds "regardless of what is causing the LLM to malfunction", the LLM being the AI model, so a fooled model and a mistaken one land in the same place. Its agentic list opens with Agent Goal Hijack, Tool Misuse and Identity & Privilege Abuse.

| Risk | What it looks like | What limits it | What Verosek does |
| --- | --- | --- | --- |
| A fooled model | A ticket or a web page tells the model to do something else, and it tries. | Limits outside the model, and a scan of what it reads. | Decides each tool call against the key's rules before it runs. With Shield on, it also reads the text. |
| Too much access | An agent that reports on orders holds a key that can delete them. | Only the tools and limits its job needs. | Offers the agent only the tools its key carries, and refuses a call over a limit. |
| The wrong customer | The agent acts on an account someone named in the chat. | An identity from your backend, not from the conversation. | Holds each tool call to the identity your backend gives it, once you set that on the key. |
| A poisoned tool | A tool's description, or a later change to it, carries instructions for the model. | Approve each definition, and review each change. | For a tool Verosek runs, the model gets only the definition your admin approved, by connecting the server or tool by tool, and a definition that changes goes back for review. |
| Too much, too fast | A loop repeats an action all night. | A cap on pace and spend, and your own check on totals. | A key's rate limit caps its pace and its budget caps model spend. Neither counts refunds, but a rule can refuse, hold or flag the same money-moving call repeated in a window. |
| No proof | Nobody can show what the agent did. | A record kept outside the agent, where a change shows. | Seals a record of each request and, by default, refuses a request it cannot record. |

*Table 1. The risks of an AI agent that can change things, and what limits each.*

## How do you secure an AI agent?

**Four controls do most of the work, and each one sits outside the model, where a persuaded model cannot reach it.**

1. **Give each agent only what its job needs.** Each agent gets its own key, and the key sets what it may do: which tools, within what limits. Verosek does not offer a tool the key does not carry, and refuses a call for one. For a tool your own code runs, give the key an allow list. *AI agent permissions: what one key should allow* goes through it.
2. **Decide each call before it runs.** Your rules read the call itself, so the conversation does not change the answer, and a blocked call never reaches your system. *Stop a tool call before it runs* follows one call, and *Restrict what SQL an AI agent can run* does the same for a database.
3. **Read what it reads.** With Shield on, Verosek reads prompts, replies, tool arguments and tool results, and the settings on that key decide what happens to a finding. No scanner catches everything, so this one comes after the first two. *What prompt injection is, and what stops it* says why.
4. **Keep a record you can check.** Verosek seals a record of each request, with secrets and personal data scrubbed before it is stored. You can ask Verosek to verify a seal, and on a self-hosted deployment an auditor holding the verification key can check a record offline. *A tamper-evident audit trail* has the detail.

*Figure 2. Where each control sits, between your agent and what it reaches. Illustrative.*

A key can also hold each tool call to the customer your backend names, whatever the model wrote. *Agent identity: whose account it acts on* covers that.

## Which framework should you use for AI agent security?

**Use a framework to decide what to check and a control to stop the call, because a list of risks refuses nothing.** OWASP describes its agentic list as "a globally peer-reviewed framework" that gives teams "a clear starting point for reducing agentic AI risks". NIST's [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is wider: NIST says it "is intended for voluntary use", and that it "is being revised". Take your questions from them. Then take each row of Table 1 and ask what would stop it on your agent today.

## Which AI security tools do you need?

**That depends on what your agents can change.** If they only call models, a router may be enough. A firewall or guardrail reads words, which suits a chat product open to strangers. An MCP gateway decides the tool call, which suits agents that write to production. They need not be three products: Verosek does all three on one key, and with Shield on, a security check is not a second vendor to buy, review, pay and log in to. *AI gateway, LLM router, MCP gateway, AI firewall* sets them side by side, and *MCP security: the risks, and what to do about each* takes MCP risk by risk.

If your own agents write to production databases and APIs, look for rules that read each tool call, a refusal when a rule cannot run, and a record you can check. Verosek gives you all three. To go through your agents with us, [book a call](https://cal.com/verosek-solutions-u9ztip/30min).

## Where does this stop?

**Verosek decides the calls it sees, and an allowed call can still be a bad idea.**

- A fooled agent can still do whatever its key allows, so keep each key narrow.
- A rule reads the call, not your records. It cannot check that a refund goes to the card that paid.
- A key's rate limit caps its pace and its budget caps model spend. Neither counts refunds, so have your own system check the total.
- Shield is off until you turn it on, and no scanner catches everything.
- Verosek does not sandbox the servers it runs. On the Enterprise plan, where the gateway runs in your environment, isolate them yourself.
- An agent that reaches a system some other way is outside all of this, because Verosek sees only what passes through it.

*Figure 3. What changes for a reviewer, and what stays yours.*

## Frequently asked questions

### Is AI agent security different from AI security?

It is the part of AI security about what an agent does. Filters and model testing ask what a model says. Agent security also asks what it may touch, who it acts for and what it did.

### How do you stop an agent that is going wrong?

Block its key, and every instance refuses that key from the next request. To stop an agent you run on Verosek, archive it so no new run starts, and cancel any background task it has going, which is best effort.

### Should a person approve every action?

No. With Verosek, a rule answers most calls at once, and a named person decides those you choose, which do not run until then.

### What happens if a check cannot run?

By default a blocking rule that cannot run refuses. A check that only flags lets the text through and names it in a response header. A request the trail cannot record is refused before anything runs.

### What does it take to put Verosek in place?

An app keeps its OpenAI, Anthropic or Gemini SDK and changes two values: the base URL and the key. The same key routes its model calls, with fallback for chat, and carries its budget and rate limits. For an MCP client, it takes two values: the endpoint and the key.
