The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions repeatedly direct the agent to run npx -y @nomadamas/k-skill@0. This command fetches and executes code from a third-party npm package at runtime, creating a significant remote code execution vector from an untrusted source.
[EXTERNAL_DOWNLOADS]: The skill depends on a third-party CLI tool and a proxy service (k-skill-proxy.nomadamas.org) that are not managed by trusted or well-known providers. This introduces supply chain risks as the agent is instructed to fetch instructions and helper files dynamically.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external food safety APIs, which creates a surface for indirect prompt injection if the source data contains malicious instructions.
Ingestion points: Data is fetched from multiple food safety endpoints via a proxy service in scripts/mfds_food_safety.py (e.g., /v1/mfds/food-safety/search).
Boundary markers: The instruction.md file includes a mandatory interview workflow and a response policy designed to prevent the agent from making medical diagnoses, though these markers may not prevent adversarial injection.
Capability inventory: The helper script (scripts/mfds_food_safety.py) is limited to network GET requests and stdout; it does not have the capability to write files or execute subprocesses internally.
Sanitization: The script includes a summarize_text function that removes HTML tags and unescapes HTML entities, providing basic sanitization of the fetched data.

mfds-food-safety