The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a sandboxExec function that allows the agent to run arbitrary Python and Shell commands. This is a high-privilege capability that could be misused if not properly constrained by the host environment.
[EXTERNAL_DOWNLOADS]: The instructions provide examples of downloading data from external sources, including api.exchangerate-api.com, google-play-scraper, and generic web APIs using urllib and pandas.
[REMOTE_CODE_EXECUTION]: The skill demonstrates patterns where data retrieved from the network is piped into a Python interpreter (e.g., curl ... | python3). Although the provided example uses a static script to process the data, the pattern itself represents a risk if user-controlled or untrusted URLs are used.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it fetches and processes untrusted external data—such as app reviews or web content—without explicit sanitization or boundary markers to prevent the data from influencing the agent's behavior.
Ingestion points: Data is ingested via google-play-scraper, urllib.request.urlopen, and pandas.read_csv in SKILL.md.
Boundary markers: No boundary markers or 'ignore' instructions are used when interpolating external data into the agent's context.
Capability inventory: The skill possesses full shell and Python execution capabilities via sandboxExec and the ability to install arbitrary pip packages.
Sanitization: External data is parsed as JSON or CSV but is not filtered for malicious natural language instructions.

sandbox-executor