The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (SAFE): The script scripts/sentiment_tools.py downloads a sentiment analysis model (uer/roberta-base-finetuned-chinanews-chinese) from Hugging Face if it is not available in the local cache. This is standard behavior for the transformers library and appropriate for the skill's stated purpose.
[COMMAND_EXECUTION] (SAFE): Database interactions in scripts/sentiment_tools.py utilize parameterized SQL queries (e.g., cursor.execute("UPDATE ... SET ... WHERE id = ?", (...))). This effectively prevents SQL injection attacks.
[CREDENTIALS_UNSAFE] (SAFE): The scripts/llm/factory.py file retrieves API keys for providers like DeepSeek, DashScope, and OpenRouter using os.getenv(). No hardcoded credentials or secrets were identified in the source code.
[DATA_EXFILTRATION] (SAFE): The skill communicates with legitimate LLM provider endpoints (OpenAI, Aliyun, etc.) and Hugging Face. There is no evidence of unauthorized data transfer or access to sensitive local files such as SSH keys or environment configuration files.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill ingests untrusted news text from a database for analysis.
Ingestion points: scripts/sentiment_tools.py via db.get_daily_news().
Boundary markers: Not explicitly used in the prompt template in SKILL.md.
Capability inventory: Subprocess calls (via BERT) and database write access (update_single_news_sentiment).
Sanitization: None; text is passed directly to the models. However, the risk is low as the model output is constrained to a sentiment score and label.

alphaear-sentiment