LM Studio develops Auto Review and Shell Judge to secure AI coding agent commands

Addressing AI Command Security

LM Studio has developed Auto Review and Shell Judge to secure commands executed by its Bionic AI coding agent. The primary goal is to prevent potentially harmful commands, such as those that could write to the filesystem, from being run when variables are involved. This system aims to identify and mitigate risks before command execution.

How Shell Judge Works

The Shell Judge component analyzes the structure of shell commands rather than relying on string-based checks, which can be insufficient due to variables and nested commands. It uses the mvdan/sh parser for Bash, Zsh, and SH, and PowerShell's own AST support. The Judge determines a command's "capabilities," identifying what it could read or change, and tracks variable values across chained commands.

Auto Review's Efficiency

Auto Review serves as the initial layer of defense, clearing a significant portion of Bionic's commands without needing to consult another language model. LM Studio reported that this layer cleared as many as 82% of commands, though this figure is considered anecdotal rather than a formal benchmark. This efficiency reduces the computational overhead of security checks.

Extensive Testing for CLI Nuances

To account for the varied behaviors of command-line tools, LM Studio has built 11,651 test cases. These tests address how different tools interpret arguments and flags, as parsing shell syntax alone is insufficient. The Shell Judge must understand how individual tools will process the input following the command to accurately assess potential risks.