BeyondTrust launches security tools to monitor AI agent actions

The rise of autonomous agents

BeyondTrust recently released a new security feature called AI Agent Security on its Pathfinder Platform. This tool is designed to manage how AI agents behave on enterprise devices, specifically targeting Windows and macOS systems. It marks a shift in security from monitoring what humans do to monitoring what non-human identities can do.

The launch follows research from BeyondTrust Phantom Labs regarding the behavior of AI coding agents. The study found that these agents often perform a high volume of downstream actions based on a single user interaction. For example, researchers observed about 1.28 million shell commands requested from roughly 132,000 prompts over a four-week period.

This data suggests that a single prompt can trigger nearly 10 commands from an AI agent. Because these actions happen at machine speed, they can quickly exceed the scope of what a human user intended. The new security layer aims to provide a way to govern these actions before they are actually carried out.

How the security layer works

The AI Agent Security tool works by identifying agents on endpoints and distinguishing their activity from human activity. It captures specific context for every action, including the human owner, the device being used, the active session, and the specific skills or Model Context Protocol (MCP) activity involved. MCP servers are tools that allow agents to connect to and interact with external services.

Once the system identifies an action, it evaluates that specific move against a set of pre-defined policies at runtime. This is different from traditional security methods that might only block known bad commands or try to detect risky behavior after it has already happened. Instead, the platform tries to determine if the agent is permitted to perform the action in the first place.

If an agent attempts an action that requires higher privileges than it currently has, the system can intervene. It can trigger a request for just-in-time access or require a human to provide manual approval. This allows the agent to complete its task while ensuring that the organization maintains oversight of sensitive permissions.

The risks of delegated access

A major concern for security teams is that AI agents often inherit the same access levels as the humans who use them. When an agent acts on behalf of an employee, it can exercise privilege across multiple platforms using that user's delegated authority. This expands the surface area for potential security breaches because the agent can operate autonomously.

The research highlighted that MCP servers are a growing area of concern because they can widen an agent's reach. Among the coding-agent users studied, 85% had at least one MCP server invoked, with some users connected to as many as 36 different servers. These connections can receive prompts and influence how the agent behaves based on the responses it receives from those external services.

Furthermore, the study found that many agent skills are bundled into code repositories. In one mid-sized enterprise, 60% of AI-active devices ran a coding agent, and users had dozens of these skills installed. Because these skills often arrive in bundles, it can be difficult for organizations to audit which ones are active or how they might change the agent's behavior.

Shifting from access to trust

The release of these tools reflects a broader shift in how enterprise technology teams view security. Because agents can make decisions and invoke tools rather than just returning information, the security equation is changing. Experts suggest that the focus is moving from simply managing who has access to continuously managing trust in real time.

Marc Maiffret, Chief Technology Officer at BeyondTrust, noted that organizations need to control privilege before an action is taken because agents operate at machine speed. He explained that the goal is to separate what a human is permitted to do from what their AI agent is permitted to do. This allows companies to use the power of AI without assuming that every human permission should automatically apply to an autonomous agent.

David Morimanno, Field Chief Technology Officer at Xalient, stated that these tools help bring trust decisions closer to the endpoint. By enforcing these rules in real time, organizations can govern AI-driven actions without slowing down the adoption of the technology. This approach helps ensure that agents only exercise the specific authority they need for a given context.