The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The file cli-wrapper.sh contains a hardcoded absolute file path C:/Users/rohit/scoop/shims/fd.exe. This exposure of a specific local username ('rohit') and the host's directory structure constitutes sensitive data exposure.
[METADATA_POISONING]: The README.md file includes extensive documentation and activation logic for unrelated skills, specifically lark-agent and lark-agent-simple. This misleading information can cause the agent to misinterpret its primary function or activate this skill for inappropriate user requests, leading to potential workflow errors.
[COMMAND_EXECUTION]: The skill is designed to substitute traditional CLI commands with modern alternatives by executing bash scripts and binaries. This grants the agent direct capability to run commands on the system, which is a significant privilege that requires clear intent and security boundaries.
[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting user-provided command arguments and interpolating them into shell execution strings within cli-wrapper.sh. While the script uses standard shell quoting, the lack of explicit sanitization for these external inputs creates a surface for indirect prompt injection or command manipulation.
Ingestion points: User messages containing keywords like 'cat', 'ls', and 'find' as defined in skill.md.
Boundary markers: No specific delimiters or warnings to ignore embedded instructions are present in the scripts or instructions.
Capability inventory: The skill allows for the execution of various bash commands including bat, eza, fd, and watchexec.
Sanitization: There is no evidence of validation or filtering applied to user-supplied command arguments before they are passed to the underlying tools.

cli-modern-tools