The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill utilizes browser_run_code and browser_evaluate to execute dynamic JavaScript snippets within the browser context. This execution occurs on pages where the user is encouraged to maintain active login sessions for services like Gemini, ChatGPT, and Claude, potentially exposing session data or personal history.\n- [COMMAND_EXECUTION]: The skill employs the Bash tool to run commands such as pkill -f "mcp-chrome". While intended for browser recovery, the use of pkill allows the agent to terminate processes on the host system.\n- [PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection (Category 8) due to its core function of aggregating data from external AI services and the web.\n
Ingestion points: Content is retrieved from multiple external AI providers and websites via Playwright extraction scripts in references/extraction.md, references/gemini.md, references/chatgpt.md, references/claude-web.md, and references/grok.md.\n
Boundary markers: The instructions do not define clear delimiters or use 'ignore embedded instructions' directives when processing or summarizing the gathered information.\n
Capability inventory: The agent possesses high-impact tools including browser_run_code for arbitrary JS execution, Bash command execution, and file system Write access.\n
Sanitization: No validation or sanitization is performed on the data extracted from the web/AI services before it is saved to local files or integrated into the final report.

ultra-research