AI agent security is the work of making sure an agent can only do what you intended — even when it reads a malicious email, misunderstands a request, or is simply wrong. Traditional application security assumes code does what it was written to do. Agents decide what to do at runtime, using tools that touch real systems. That changes where the risk sits and which controls actually help.

Why AI agent security is different

A chatbot's worst outcome is usually a bad answer. An agent's worst outcome is a bad action: a refund that shouldn't have happened, customer data sent to the wrong place, a production resource deleted. Three things make agents harder to secure than ordinary software:

  • They take instructions from data. Anything an agent reads — a web page, a support ticket, a document — can contain text that changes its behaviour. This is prompt injection, and there is no complete fix at the model level.
  • They act with delegated authority. Agents call tools with API keys, service accounts or a user's permissions. If the agent is manipulated, it uses that access on the attacker's behalf.
  • Their behaviour isn't fixed. The same agent can take different paths for similar requests, and a model upgrade can change behaviour overnight.

Security teams are noticing. In a 2026 Cloud Security Alliance study commissioned by Zenity, only 8% of organisations said their AI agents never exceed their intended permissions.

The main risks

OWASP now publishes two lists that matter here. The OWASP Top 10 for LLM Applications ranks prompt injection first and includes excessive agency — agents with more functionality, permissions or autonomy than they need. The newer OWASP Top 10 for Agentic Applications (published December 2025) starts with agent goal hijack, tool misuse and identity and privilege abuse. In plain terms, the risks you need to plan for are:

  1. Goal hijack — injected instructions redirect what the agent is trying to do.
  2. Tool misuse — the agent uses a tool it's allowed to use, in a way nobody intended.
  3. Over-privilege — the agent holds broader access than its job needs. See excessive agency.
  4. Data exposure — the agent reads or sends data it shouldn't.
  5. Unaccountable actions — nobody can reconstruct what the agent did, or who allowed it.

Controls that actually work

The common thread: don't rely on the model to police itself. Put controls where the agent touches the world.

Limit what the agent can reach

Give each agent only the tools and scopes its job requires — the least privilege principle applied to agents. A support agent that can look up orders doesn't need to delete them.

Check every action before it runs

Put a policy check between the agent and its tools. Each tool call is evaluated against rules — amount, data type, environment, customer — and is allowed, denied, or held for a person. This is the core of AI agent guardrails that can't be talked past.

Keep a person in the loop for high-impact actions

Route actions that are hard to reverse to a named approver. Done well, human in the loop adds safety without slowing everything down.

Record every decision

Keep an AI agent audit trail: what was attempted, what was decided, by which rule, and who approved it.

Treat tool connectors as part of your attack surface

Tool servers — including MCP servers — can be compromised or misconfigured. Review them like any third-party dependency.

Where to start

You don't need a full programme on day one. A practical first month:

  1. List every agent and the tools each one can call.
  2. Mark the actions that move money, touch customer data, contact customers or change production.
  3. Put a policy check in front of those actions, with clear allow / approve / deny rules.
  4. Turn on decision logging and review it weekly.
  5. Expand coverage as you add agents and tools.

Go deeper

Where Gaten fits

Gaten is a runtime checkpoint between your agents and the tools they call. Every action is checked against your rules before it runs — allowed, denied, or sent to a person — and every decision is recorded. See how it works or compare it with other approaches.

Frequently asked questions

What is AI agent security?

It's the set of controls that keep AI agents doing only what they're supposed to: limiting their access, checking their actions before they run, involving people for high-risk decisions, and recording what happened.

Is prompt injection the biggest AI agent security risk?

It's the most cited, because it's hard to prevent at the model level. That's exactly why the strongest controls sit outside the model, at the point where the agent takes an action.

Do we need a separate security tool for agents?

Existing identity, network and logging tools still matter. What they usually don't do is evaluate each agent action in context — the amount, the data, the customer — before it runs. That's the gap runtime policy fills.