The Agent Skills Directory

[COMMAND_EXECUTION]: The skill mandates the use of dangerouslyDisableSandbox=true when executing the Bash tool to run the glm-review CLI. This instruction explicitly directs the agent to bypass built-in security protections and sandboxing mechanisms, potentially granting the external command unrestricted access to the host system.
[COMMAND_EXECUTION]: The skill requires the execution of an unverified third-party CLI tool (glm-review). This tool is not part of a standard system distribution or a recognized trusted library, and its internal logic is opaque to the static analysis.
[DATA_EXFILTRATION]: The skill uses git show and git diff to extract source code changes (including newly created files) and writes them to a temporary file (/tmp/glm-review-diff.patch). This data, which may contain proprietary logic or sensitive information, is then passed to an external tool and its associated cloud-based model (GLM-5).
[COMMAND_EXECUTION]: The skill uses the run_in_background=true flag. While useful for long-running tasks, when combined with disabled sandboxing and unverified external code, it increases the risk of persistent or stealthy malicious activity occurring outside the agent's immediate observation.

rrr