The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external forum discussions (Google Forums, Reddit, Quora) via the Frevana API which could contain malicious instructions designed to influence the agent.
Ingestion points: Search results are retrieved from https://ai-factory.frevana.com in scripts/search_google_forums.sh and returned to the agent.
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are used to isolate external content from the agent's context.
Capability inventory: The skill possesses file-writing capabilities (cp) and network access (curl).
Sanitization: The script performs structural JSON validation but does not sanitize or filter the content within the search results.
[COMMAND_EXECUTION]: The helper script executes shell commands including curl for network communication and cp/mkdir for local file management.
Evidence: scripts/search_google_forums.sh allows the user or agent to specify an arbitrary --output path which is subsequently used in a cp command to write data.
[DYNAMIC_EXECUTION]: The bash script generates and executes Python code at runtime using heredocs to perform JSON serialization and validation.
Evidence: scripts/search_google_forums.sh contains two blocks of python3 - that process environment variables and API responses dynamically.
[DATA_EXFILTRATION]: The skill sends user-provided search queries and an authentication token to an external third-party domain.
Evidence: The script performs a POST request to https://ai-factory.frevana.com using the FREVANA_TOKEN for authorization.
[EXTERNAL_DOWNLOADS]: The skill connects to a non-whitelisted external API to retrieve search data.
Evidence: Network communication with https://ai-factory.frevana.com is required for the skill to function.

google-forums-search