The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to capture and analyze content from third-party websites, creating a surface for indirect prompt injection. Malicious instructions embedded in the scraped text could potentially influence the agent's behavior during subsequent analysis steps.\n
Ingestion points: The skill uses Playwright MCP tools like browser_navigate and browser_evaluate to fetch content from arbitrary URLs in files such as SKILL.md and templates/capture-workflow.py.\n
Boundary markers: The provided templates do not include explicit instructions or delimiters to isolate captured content from the agent's system prompt.\n
Capability inventory: The skill utilizes browser interaction tools (browser_click, browser_fill_form, browser_evaluate) and makes network requests via the httpx library.\n
Sanitization: While templates/extract-article.js removes noise elements, it does not sanitize the extracted text for malicious natural language instructions.\n- [DATA_EXFILTRATION]: The skill includes functionality to send captured data to external HTTP endpoints. While intended for RAG integration, this network activity represents a potential exfiltration vector if the target endpoints are modified for malicious purposes.\n
Evidence: The skill provides examples of sending data to remote APIs using the httpx library in SKILL.md and templates/queue-to-example.py.

browser-content-capture