The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection (Category 8) by processing external data from benchmark datasets.
Ingestion points: scripts/run_eval.py loads data from various HuggingFace datasets (e.g., pig4431/HeQ_v1, HebArabNlpProject/HebrewSentiment).
Boundary markers: Absent. Prompt templates in scripts/run_eval.py and references/prompt-templates.md use simple labels like Passage: and Text:, but lack explicit instructions to the target model to ignore any embedded commands or instruction-overrides within the data.
Capability inventory: scripts/run_eval.py contains network operation capabilities, sending the assembled prompts (containing the untrusted data) to several LLM providers (Anthropic, OpenAI, Google, and AI21).
Sanitization: Absent. The data from the external datasets is interpolated directly into prompt strings without validation or escaping.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of external components and data.
Package Installation: Instructions in SKILL.md require installing datasets, transformers, anthropic, openai, google-genai, and ai21. These are well-known, legitimate libraries from standard registries.
Data Downloads: scripts/run_eval.py uses the datasets library to fetch content from HuggingFace. The sources (pig4431, HebArabNlpProject) are associated with the benchmark datasets cited in the documentation.

hebrew-llm-eval-suite