The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes an LLM-as-judge evaluation pattern that is susceptible to indirect prompt injection. 1. Ingestion points: Data from model predictions (inputs and outputs) is passed to a judge model in the llm_judge_sample function within SKILL.md. 2. Boundary markers: The JUDGE_PROMPT example lacks delimiters or specific instructions to isolate untrusted data from the judge's instructions. 3. Capability inventory: The skill utilizes Write and Bash tools and provides code for network requests via libraries like httpx and the anthropic SDK. 4. Sanitization: There is no evidence of input validation or sanitization in the provided evaluation code snippets.
[COMMAND_EXECUTION]: The skill uses the Bash tool to perform project tasks like npm test and npm run build. These commands are standard for the described evaluation and build processes.
[EXTERNAL_DOWNLOADS]: The skill contains Python code for interacting with external APIs, such as the Anthropic API and model endpoints, for the purpose of shadow testing and response grading. These network operations are consistent with the documented MLOps use cases.

eval-harness