The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The sentiment analysis functionality in scripts/sentiment_tools.py is vulnerable to indirect prompt injection. Untrusted content fetched from the web is directly interpolated into a prompt for the LLM without any boundary markers or sanitization, which could allow malicious web content to influence the agent's behavior.
Ingestion points: scripts/search_tools.py fetches web content using ContentExtractor (via Jina Reader) and passes it to SentimentTools for analysis.
Boundary markers: No delimiters or "ignore embedded instructions" warnings are present in the prompt template in scripts/sentiment_tools.py (lines 117-128).
Capability inventory: The SentimentTools uses an agent to perform analysis. While not directly assigned tools in this script, it operates within a skill environment that has search, network, and database capabilities.
Sanitization: The text is truncated to 1000 characters, but no escaping or validation of the content is performed.
[EXTERNAL_DOWNLOADS]: The skill fetches external resources from well-known services for its core functionality.
Content Extraction: Uses the Jina Reader API (https://r.jina.ai/) to fetch and format web page content.
Web Search: Utilizes Jina Search (https://s.jina.ai/), DuckDuckGo, and Baidu for information retrieval.
Machine Learning Models: Automatically downloads the BERT sentiment analysis model uer/roberta-base-finetuned-chinanews-chinese from Hugging Face when configured in BERT mode.

alphaear-search