The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation includes examples using npx jev guard, which downloads and executes the jev package from the npm registry at runtime.
[COMMAND_EXECUTION]: The skill instructions provide shell commands (npx, python) to run guardrail batteries and process text, which involves executing external scripts and utilities.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data (user requests and model responses) to detect hazards, creating an inherent injection surface.
Ingestion points: Text is passed to the guardrail via the --text CLI argument and functions like inputBattery or outputBattery mentioned in SKILL.md.
Boundary markers: The provided instructions do not specify the use of delimiters or boundary markers for the data being screened.
Capability inventory: The skill has the capability to execute shell commands (npx) and run Python scripts (examples/07_guardrails.py).
Sanitization: The skill mentions redacting secrets before sending state but does not explicitly detail input sanitization or escaping mechanisms for the text being analyzed.
[PROMPT_INJECTION]: The text contains the pattern "Ignore previous instructions" in SKILL.md. Analysis confirms this is a literal string used within a code example to demonstrate how the guardrail identifies injection attempts, rather than an active attempt to subvert the agent's behavior.

jev-guardrail