The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection in the LLMAnalyzer class within scripts/llm_analyzer.py.
Ingestion points: Untrusted text is ingested through multiple methods such as analyze_sentiment_llm, extract_topics_llm, and summarize_text_llm, as well as from sample files like examples/sample_data/sample_reviews.csv.
Boundary markers: Absent. User-provided text is interpolated directly into prompts without using delimiters or boundary markers to isolate data from instructions.
Capability inventory: The skill configuration in SKILL.md allows the use of tools such as Bash, Write, and Edit.
Sanitization: No sanitization or filtering is performed on the input text before it is interpolated into LLM prompts.
[EXTERNAL_DOWNLOADS]: The skill fetches necessary NLP data files from official sources.
Evidence: scripts/text_analyzer.py and scripts/sentiment_analyzer.py use nltk.download() to retrieve lexicons and tokenizers.
[DATA_EXFILTRATION]: The skill communicates with external AI service providers.
Evidence: scripts/llm_analyzer.py performs network requests to api.openai.com and dashscope.aliyuncs.com to perform text analysis.
Note: These communications are intended for the primary purpose of the skill and target well-known service providers.

content-analysis