The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process arbitrary external text, such as user prompts, model outputs, and assertions, for evaluation purposes. This creates a vulnerability surface for indirect prompt injection where malicious instructions embedded in the processed data could influence the grader's behavior.
Ingestion points: External data is ingested via variables such as $prompt, $output_text, and $assertions_json as demonstrated in the shell patterns in references/invocation-patterns.md.
Boundary markers: The provided command templates lack strong delimiters or explicit instructions to the AI to ignore any commands embedded within the data being evaluated.
Capability inventory: The skill facilitates execution of the claude CLI, which has capabilities for file system access and tool use.
Sanitization: The analyzed snippets do not include mechanisms for sanitizing or escaping the interpolated content to prevent it from being executed as instructions.

bare-eval