The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates an attack surface for indirect prompt injection by processing external, untrusted data to perform evaluations and generate reports.
Ingestion points: The skill ingests user-provided task descriptions and responses from target AI model endpoints for comparison.
Boundary markers: No specific delimiters or boundary instructions are mentioned in the configuration to isolate untrusted model outputs from the judge's instructions.
Capability inventory: The skill performs network operations to query external APIs and uses file system access to save generated reports, charts, and checkpoints.
Sanitization: There is no evidence of sanitization or escaping of ingested model responses before they are presented to the judge LLM.
[EXTERNAL_DOWNLOADS]: The skill documentation instructs the user to install the py-openjudge and matplotlib Python packages. py-openjudge is a vendor-owned resource from the same organization that authored the skill, and matplotlib is a well-known library for generating visualization charts.

01-auto-arena