The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted content from various web sources to inform its campaign builder logic, creating a surface for indirect prompt injection.\n- Ingestion points: Scrapes competitor ads from the Meta Ad Library, market research from Reddit, and user landing pages via WebFetch.\n- Boundary markers: The instructions do not specify the use of delimiters or warnings to ignore embedded instructions in the scraped data.\n- Capability inventory: Access to agent-browser, WebSearch, WebFetch, and Gemini image generation.\n- Sanitization: External content is used to influence the final output without filtering or validation.\n- [COMMAND_EXECUTION]: The skill provides a shell command template that interpolates user-controlled variables, posing a risk of command injection.\n- Evidence: The bash script in Stage 00, Step 2 includes npx agent-browser open "https://www.facebook.com/ads/library/...&q=[competitor name or keyword]".\n- Risk: User-supplied keywords are not sanitized before being placed into the shell command, allowing for potential execution of arbitrary code if the input contains shell metacharacters.

meta-campaign