The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is specifically designed to process untrusted data sourced from external web searches and API results, which creates a potential attack surface.
Ingestion points: Untrusted external data enters the agent's context through the results array (including titles, URLs, and snippets) passed to the jev search tool as described in SKILL.md.
Boundary markers: The skill uses a structured JSON payload to separate query metadata from result content, and results are handled by unique IDs (P0, P1, etc.) to isolate content.
Capability inventory: The agent using this skill typically performs searches (web_search), reads result content, and writes research summaries based on the tool's guidance.
Sanitization: The skill documentation claims to implement a "local screen" and a "Jev ranking" process to identify prompt injection attempts (e.g., instructions like "ignore your guidelines") and mask sensitive information such as emails, phone numbers, and tokens. It explicitly instructs the agent to disregard result IDs flagged in the dropped_injection_ids output field.
[DATA_EXFILTRATION]: The skill documentation discloses that specific data (the question, queries tried, and up to 900 characters of each result snippet) is sent to an external service ("Jev") for processing.
The skill provides privacy guidelines, advising against including customer or student records and claiming to mask PII and credentials before transmission.

jev-search