AI Agent Security | AI Agent Security

What is AI Agent Security?

AI Agent Security gives security and IT teams visibility into, and control over, the AI agents running on their endpoints: AI coding tools, desktop AI apps, the MCP servers they connect to, and the skills they use.

The gap it fills is organizational as much as technical. AI agents are typically adopted by developers and other teams to improve productivity and delivery speed, often without the security team's knowledge. Leadership encourages AI adoption, but security teams have generally lacked the tooling to manage the risk that adoption creates.

The product depends on two things:

• The Pathfinder agent. The BeyondTrust agent installed on your endpoints so activity can be monitored and enforced. It is not an AI agent.
• A policy. The policy dictates what is blocked, what is audited, what needs approval, and what is allowed.

AI agents are what AI Agent Security discovers on those endpoints, such as Claude Code and OpenAI Codex.

Terminology

How is it useful to my organization?

• AI Agent Security removes the blind spot AI agents create on your endpoints. An agent runs as the signed-in user, with no identity of its own, so an unmonitored agent can take any action its user can. That reach does not stop at the endpoint: the cloud services, repositories, and business applications the user signs in to cannot tell the agent's actions from the person's.
• AI Agent Security shows you every agent running in your environment, applies policy to each agent separately from the person who runs it, and stops risky actions before they complete rather than flagging them after the fact. The result is an audit trail that separates human activity from AI activity, so your security team can adopt AI tooling without losing accountability.

Problems this addresses

• No visibility into what is deployed. Organizations frequently do not know how many AI agents exist in their environment, who installed them, where they run, or what they do. The real number is often far higher than anyone expects.
• Citizen developers. AI tooling is no longer confined to engineering. People without a security background build and use AI tools, and may introduce vulnerabilities without recognizing it.
• Reactive security. Many tools detect malicious activity and alert only after the action completes. By then the damage is done.
• Agents using a person's access. An agent working for a developer can reach the same production systems, credentials, and cloud command-line tools that developer can, such as running a cloud CLI against a production account.
• Prompt injection and poisoned prompts. A manipulated prompt can steer an agent toward an action nobody asked for. AI Agent Security does not inspect or block the prompt; policy governs the action the agent then attempts, so a blocked action stays blocked however the request was phrased.
• Unsafe package installation and supply chain risk. For example, an agent bypassing an approved package repository.

What the product gives you

• Discovery. Until the Pathfinder agent is deployed, AI agents, the MCP servers they connect to, the LLM endpoints they call, and the skills they use are invisible.
• Auditability. AI sessions record the prompts entered, the commands run, the files accessed, the network activity, and which tools ran. Events provide the same record per action.
• Visible risk. Each AI session lists its findings, the actions a policy blocked or audited, with their severity. The Dashboard surfaces the riskiest agent behavior across your fleet.
• Preventative control. Policy determines whether an action is allowed, audited, held for approval, or blocked before it executes, rather than reporting it afterward.
• Measurable impact before enforcement. Before you save a policy, you can replay recorded activity from a past time window and see what the policy would have blocked, so a rule can be checked before it reaches anyone.

How do I access AI Agent Security?

1. Sign in to app.beyondtrust.io with your credentials.
   The Home page displays.

2. Select the AI Agent Security tile.

   

Licensing and prerequisites

To add AI Agent Security to your tenant, contact your BeyondTrust account team.

Before you deploy the Pathfinder agent, confirm the following:

AI Agent Security supports Windows and macOS endpoints only.

How enforcement works

Separating agent identity from human identity

AI Agent Security distinguishes actions taken by an AI agent process from actions taken by a human, and applies different rules to each, even when both run on the same machine under the same credentials.

This distinction is the foundation for everything else. A developer can keep access to a production environment while the agent running on that developer's machine is denied it.

Agents are identified using classifiers that include process naming, signatures, and LLM API traffic patterns. Because AI agents consistently communicate with an LLM API, they are relatively straightforward to detect.

Enforcement point

Enforcement is applied on the endpoint, at the kernel and network level, rather than by screening prompts or shell commands. For AI coding tools that support hooks, policy can also act on the tool calls the agent makes.

This matters for reliability. Screening prompts, or screening shell commands and file writes, is fragile: there are many ways around it, and an AI agent is capable of finding them. Applying control below that layer means enforcement holds regardless of which agentic process is involved and regardless of how the request is phrased.

Kernel-level operation also provides deep visibility into the processes running on an endpoint and the network traffic leaving it, which is what allows the product to see what an agent is actually doing rather than inferring it from logs after the fact.

Consequences for users

• No workflow change is required. Enforcement applies whether or not the user runs their agent with permission prompts turned off, and does not depend on adopting a sandbox or opting in to a gateway or proxy.
• Coverage is not limited to the command line. Any agentic process is covered, including installed desktop applications and packaged coding agents. If a desktop AI application attempts to use a cloud provider CLI, the same policy applies as for a terminal-based agent.
• Gateway bypass does not defeat it. Because control is applied on the endpoint, MCP and network controls do not depend on the agent being routed through a particular ecosystem.