The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted data from web search results.
Ingestion points: Untrusted data enters the context through web search results and scraped HTML content, which are interpolated into the {SEARCH_RESULTS_CONTEXT} variable.
Boundary markers: The prompt template used for summarization lacks delimiters or instructions to ignore embedded commands within the fetched context.
Capability inventory: The skill utilizes requestUrl for network access and the LLM for content generation.
Sanitization: No evidence of content sanitization or instruction filtering is mentioned in the documentation or logic flow.
[DATA_EXFILTRATION]: Risk of credential exposure via debug logs.
Evidence: The enableApiErrorDebugMode configuration logs the "Full API request/response." If the tavilyApiKey is transmitted via headers or query strings, it could be inadvertently exposed in the agent's logs or console output.
[EXTERNAL_DOWNLOADS]: The skill downloads content from arbitrary external URLs.
Evidence: The DuckDuckGoProvider performs web scraping and fetches full webpage content using requestUrl. While this is a core feature for research, it involves interacting with untrusted external servers and processing their data.

notemdpro-web-researcher