The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, such as agent entry points, tool definitions, usage documentation, and ATIF traces, to generate evaluation cases and grading criteria.
Ingestion points: Data enters the context during Stage 4 (Understanding the evaluation starting point) and Stage 5 (Prepare cases and grading) when reading repository files and agent documentation mentioned in SKILL.md.
Boundary markers: The instructions do not specify explicit delimiters or 'ignore embedded instructions' markers when interpolating untrusted documentation into the case planning process.
Capability inventory: The skill utilizes Bash, Write, Read, Grep, and Glob tools across all stages to create and execute evaluation tasks.
Sanitization: No explicit sanitization or validation of the ingested natural language documentation is mentioned before it is used to influence the generation of evaluation logic.
[DYNAMIC_EXECUTION]: The skill generates executable content at runtime, specifically NeMo Gym environment modules and Harbor task scripts (test.sh, solve.sh).
Evidence: The skill uses nemo_gym.environment.scaffold.scaffold_environment to create Python modules and populates Harbor templates as described in references/gym-first-eval.md and references/harbor-first-eval.md.
Severity Context: This is categorized as low severity because the code generation follows established templates for evaluation purposes.
[COMMAND_EXECUTION]: The skill executes various system-level commands through the Bash tool to manage evaluations and environments.
Evidence: Instructions include running gym env validate, gym env test, harbor trial start, harbor job start, and docker image rm to validate and run evaluations as documented in the reference files.

eval-author-first-eval