The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by facilitating the ingestion of external, untrusted data that is then processed within a high-privilege cloud environment. 1. Ingestion points: Data enters the system through FastAPI endpoints (@modal.fastapi_endpoint), CSV file processing (pd.read_csv), and external API calls (requests.get). 2. Boundary markers: The provided reference implementation snippets do not demonstrate the use of clear delimiters or instructions to ignore instructions embedded within external data. 3. Capability inventory: The skill provides access to shell environments (Bash), network resources, and the ability to write to persistent storage. 4. Sanitization: There are no explicit examples of input sanitization or validation shown for the data ingested from remote sources.
[COMMAND_EXECUTION]: The skill instructions include examples of shell command usage and subprocess management. The documentation proactively provides security guidance, recommending the use of list-based arguments for subprocess.Popen to mitigate command injection risks.
[EXTERNAL_DOWNLOADS]: The skill manages the installation of various Python dependencies and the retrieval of external assets, such as cloning repositories from github.com/modal-labs. These actions are standard for the tool and utilize well-known, legitimate service providers.
[DYNAMIC_EXECUTION]: The platform's core functionality involves the dynamic deployment and execution of Python code in serverless containers, including build-time execution of functions to prepare environments.

alterlab-modal