The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs users to install the officecli tool by piping scripts from a remote URL directly into the system shell. This practice is inherently risky as it executes remote code with the user's privileges without prior inspection.
Evidence in SKILL.md: curl -fsSL https://d.officecli.ai/install.sh | bash for macOS/Linux systems.
Evidence in SKILL.md: irm https://d.officecli.ai/install.ps1 | iex for Windows (PowerShell).
[COMMAND_EXECUTION]: The skill uses the officecli tool to build document components via shell commands. While the instructions include guidance on shell quoting, the interpolation of user-supplied text into these commands poses a risk of command injection if input is not properly sanitized.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it ingests untrusted user data, such as research paper content and abstracts, and uses it as arguments for command-line tools.
Ingestion points: User-provided text for titles, author names, abstracts, and body content passed through officecli add commands.
Boundary markers: The skill suggests using single quotes for values containing characters like $, but lacks comprehensive delimiters for all untrusted input.
Capability inventory: The agent can execute officecli commands and manage file system objects (specifically .docx files).
Sanitization: No explicit sanitization or validation of the user-provided text is implemented before it is processed by the tool.

officecli-academic-paper