The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The setup instructions direct the user to download and execute shell scripts from a remote domain (d.officecli.ai) using patterns like curl | bash and irm | iex. While these represent the vendor's official installation method, they are inherently high-risk patterns.
Evidence: curl -fsSL https://d.officecli.ai/install.sh | bash and irm https://d.officecli.ai/install.ps1 | iex in SKILL.md.
[DYNAMIC_EXECUTION]: The skill uses python3 -c to perform calculations on variables that are directly interpolated from shell environment variables. These variables ($STATED, $PLUGGED) contain data read from external Excel cells. If these cells contain malicious content, they could potentially execute arbitrary Python code during the audit gates.
Evidence: `python3 -c "print(abs(float('$STATED')
float('$PLUGGED')))"inSKILL.md`.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from Excel files (.xlsx) and uses the results to drive control flow decisions and audit checks without explicit sanitization.
Ingestion points: Data enters the context via officecli get and officecli query commands executed on the target Excel file in SKILL.md.
Boundary markers: No delimiters or instructions to ignore embedded commands are present when processing cell values.
Capability inventory: The skill uses officecli (file read/write), bash (shell execution), and python3 (dynamic code evaluation).
Sanitization: There is no evidence of validation or escaping for cell values before they are used in shell variables or Python commands.
[COMMAND_EXECUTION]: The skill relies heavily on shell piping and command substitution to perform data validation and audit gates, which increases the attack surface for injection if any input sources are compromised.
Evidence: Extensive use of backticks and $() for command substitution throughout the SKILL.md build gates.

officecli-financial-model