The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill contains several templates that utilize the curl | bash pattern for software installation. Although the URL https://example.com/install.sh is a placeholder, providing these patterns encourages the use of insecure installation methods that bypass integrity checks.
[COMMAND_EXECUTION]: Multiple documentation templates include shell commands for installation (apt install, brew install, make install) and initialization (mytool init). If an autonomous agent attempts to "verify" the documentation it generates, it might attempt to execute these commands.
[PRIVILEGE_ESCALATION]: The "Shell Completions" template includes a command to write to /etc/bash_completion.d/, a system directory that typically requires root or administrative privileges.
[INDIRECT_PROMPT_INJECTION]: As a tool designed to process project-specific information into documentation, the skill has an ingestion surface for potentially untrusted data.
Ingestion points: User-provided project details and descriptions used to fill the templates (found in SKILL.md).
Boundary markers: The skill uses markdown code blocks to delimit generated content, which provides some separation.
Capability inventory: The skill itself primarily generates text but exists within an environment where the agent may have file-writing or command-execution capabilities.
Sanitization: No explicit sanitization or validation of the input data is mentioned in the instructions.

readme-craft