The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill handles untrusted web data retrieved via the Exa.ai Answer and Search APIs, which is a known vector for indirect prompt injection where instructions embedded in web pages could influence agent behavior.\n
Ingestion points: Web content and AI-generated responses retrieved through exa.answer(), exa.search_and_contents(), and exa.research() in references/answer-api.md and references/deep-search.md.\n
Boundary markers: None found. The provided code snippets do not demonstrate the use of delimiters or 'ignore' instructions for external data.\n
Capability inventory: All capabilities are limited to data retrieval, string formatting, and async task management; no dangerous execution sinks like eval(), exec(), or subprocess are present.\n
Sanitization: Examples in references/citations.md provide basic URL validation and keyword overlap checks, but they lack robust sanitization for detecting malicious natural language instructions.\n- [External Dependencies] (LOW): The skill references the exa-py and exa-js libraries in its code snippets. These are the official SDKs for the Exa.ai service and are standard for this use case.

exa-research