The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external, untrusted data including model methodology documents, datasets, and label definitions. This creates a surface for indirect prompt injection where malicious instructions embedded in the audited data could influence the agent's behavior during the audit pipeline.
Ingestion points: Data reconstruction (SKILL.md), Target / Label Analysis (SKILL.md), and Feature Analysis (SKILL.md) sections involve processing external data structures.
Boundary markers: No explicit instructions are provided to the agent to treat audited data strictly as content and ignore potential instructions within it.
Capability inventory: The skill has the capability to write files and execute code via the shell (SKILL.md, run.sh), which increases the potential impact of an injection.
Sanitization: There are no documented steps for sanitizing or filtering instructions from the external data.
[DYNAMIC_EXECUTION]: The skill's core functionality requires the dynamic generation and execution of scripts to replicate model training pipelines and perform statistical tests.
Evidence: The skill includes a 'Reproducibility Standard' requiring that 'Every analysis must be fully reproducible' and 'Scripts must be versioned and self-contained' (SKILL.md). It also provides several Python templates for metrics and interpretability that are intended to be executed at runtime.
Execution Mechanism: The agent is instructed to 'Reproduce model training pipeline from documented specifications', which involves generating and running code based on external inputs.

specialized-model-qa