BeyondTrust says each AI instruction sends almost 10 commands
BeyondTrust says 85% of the coding agent users it saw had an MCP server, a connection that lets an agent work on other platforms.
Claimed, not confirmed
On 21 October 2026, BeyondTrust released AI Agent Security, a tool that controls the tasks that AI agents can do on Windows and macOS computers. The company saw about 1.28 million shell commands, which are text orders to a computer, from about 132,000 instructions to AI in four weeks. The data is from samples of customers, and no other group checked it.
How it works
The tool tells agent work from person work. It checks each agent step against policy before the step operates. It can use just-in-time or person approval when an agent must have more privilege. It does not limit the tasks that employees can do.
The numbers
- Almost 10 shell commands for each instruction.
- 85% of coding agent users had one or more MCP servers.
- 4.35 MCP servers for each user, and 36 for one user.
- About half of agent skill installs came in code repositories.
What is not known
BeyondTrust gave no test of how the tool stops attacks. It did not tell how many customers are in the samples, or if the numbers are the same in other areas.
Insight: Vendors count agent steps, then sell the tool that counts
Mintlify says 83% of the companies in its report let agents write documentation updates. In one test, a rule monitor gave an alarm for 29.3% of safe agent runs. BeyondTrust counts agent work and sells the control, which reads as a move to be the place where agent steps are checked. Get the alarm rate from each vendor.
Sources
Posted
