The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes a research phase that ingests data from third-party sources. It addresses this attack surface with a comprehensive 'Untrusted content boundary' contract that mandates the use of <EXTERNAL_DATA> delimiters and instructs the agent to treat retrieved content as data rather than instructions, which effectively manages the risk of malicious instructions embedded in web content.
[COMMAND_EXECUTION]: The skill uses local Node.js and Bash scripts to run a mockup server for UI designs. The server implementation follows security best practices by binding to the loopback interface by default, validating session tokens with timing-safe comparisons, and performing path validation to prevent directory traversal.
[DATA_EXFILTRATION]: The instructions explicitly prohibit the transmission of sensitive context or credentials to third parties. The server components are configured to stay within the local environment, and user events from the browser are sanitized before being recorded for agent processing.

brainstorming