The Agent Skills Directory

[DATA_EXFILTRATION]: The skill documentation includes a hardcoded absolute file path for a virtual environment. \n
Evidence: source /Users/zhangmingda/clawd/.venv/bin/activate found in SKILL.md. \n
This leaks the local system username (zhangmingda) and directory layout to the agent and potential skill consumers.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process and analyze untrusted text data provided by users or extracted from documents. \n
Ingestion points: The preprocess(text) function and model fitting functions (e.g., tfidf.fit_transform(documents), topic_model.fit_transform(documents)) process external text corpora. \n
Boundary markers: No explicit delimiters or instructions to ignore embedded commands were observed in the code snippets. \n
Capability inventory: The skill primarily performs text analysis, classification, and extraction. It does not possess direct file-writing or network-exfiltration capabilities beyond standard library model downloads. \n
Sanitization: Employs basic regex cleaning (re.sub(r'[^\w\s]', '', text)) and NLTK-based tokenization, which provides minimal protection against sophisticated prompt injection.\n- [EXTERNAL_DOWNLOADS]: The skill fetches linguistic resources and pre-trained models from established external sources. \n
Downloads NLTK data (such as punkt, stopwords, and wordnet) from official NLTK mirrors. \n
Downloads pre-trained models from Hugging Face official repositories using the transformers and sentence-transformers libraries (e.g., all-MiniLM-L6-v2). \n
References third-party packages from the standard Python Package Index (PyPI) (e.g., bertopic, jieba).

nlp-analysis