The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses behavioral instructions in its documentation to constrain agent actions, specifically requiring explicit user approval before performing any database writes. While these are safety-oriented constraints, they represent instructions designed to override default agent behavior.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its LinkedIn content analysis feature. Scraped post text is interpolated directly into an LLM prompt for scoring without the use of boundary markers or delimiters.
Ingestion points: Scraped LinkedIn posts retrieved from Apify in scripts/signal_scanner.py.
Boundary markers: None present; content is directly inserted into the prompt template in the _llm_score_post function.
Capability inventory: The resulting score can trigger database inserts into the signals table and status updates in the people table.
Sanitization: No sanitization or validation is performed on the untrusted external content before it is processed by the LLM.
[DATA_EXFILTRATION]: The skill accesses sensitive credentials, including SUPABASE_SERVICE_ROLE_KEY and APIFY_TOKEN, from the environment. While these are necessary for the skill's defined functionality of scanning and updating lead data, they represent a significant access point to the user's data infrastructure.

signal-scanner