Skip to content
StrikeCyberStrikeCyber
Field Manual · Concept

AI Agent Security

Reviewed November 13, 2026Concept 2 min read

Quick answer

AI agent security is the practice of protecting AI systems that can take actions on their own, such as reading email, browsing the web, running code, calling APIs or changing records. Unlike a chatbot, an agent turns a manipulated instruction into a real action, so the impact of an attack is limited only by the tools and permissions the agent holds.

On this page
01

How It Works

An agent combines a language model with tools (often connected through the Model Context Protocol), memory and the user's or a service account's permissions. It plans steps and executes them, often reading untrusted content along the way: web pages, emails, documents and tool descriptions. Any of that content can carry indirect prompt injection that redirects the agent. Attackers also target the tools themselves through tool poisoning, and the agent's own over-broad permissions create excessive agency.

02

Why It Matters

Organizations are connecting agents to email, files, code repositories, ticketing and finance systems faster than they are testing them. A single malicious email or web page that an agent reads can lead to data being sent to an attacker, unauthorized changes or actions taken under a trusted identity, with logs that show only that the agent did it.

03

How to Test for It

  1. Inventory every agent, the tools it can call, the data it can read and the identity it acts as.
  2. Test whether instructions hidden in emails, documents, web pages or tool outputs can change the agent's actions.
  3. Test whether the agent can be made to send data to an external destination, such as through links, images or API calls.
  4. Check whether high-impact actions require human confirmation and whether that confirmation can be bypassed.
  5. Review logs to confirm you can reconstruct what the agent read, decided and did.
04

How to Reduce the Risk

Give each agent the narrowest tools and permissions it needs, use separate identities for agents, require human approval for high-impact or irreversible actions, treat all retrieved content as untrusted, restrict outbound destinations, vet and pin the tools and MCP servers agents use, log every tool call, and red team agents before deployment and after every significant change.

05

Regulatory and Framework Context

In the US, the NIST AI Risk Management Framework and OWASP's guidance on agentic AI threats address autonomous actions, and CISA's AI guidance stresses least privilege and human oversight for AI systems with access to sensitive resources.

Being exploited now

Threat Radar →

The latest related vulnerabilities confirmed as exploited in the wild.

Latest intel on this topic

Updated hourly from the daily briefing. More in ai security news.

Common questions

How is agent security different from chatbot security?

A chatbot can only say something wrong. An agent can do something wrong, such as send data, change records or run code, so the same manipulation has a far bigger impact.

What is the biggest risk with AI agents today?

Indirect prompt injection combined with excessive permissions: untrusted content the agent reads can redirect it to misuse the access it already has.

Can we test our AI agents ourselves?

Basic checks, yes. A thorough assessment needs testers who combine prompt injection techniques with knowledge of the connected systems, which is what AI red teaming covers.

Change log

  1. November 13, 2026First published.
(877) 657-8496