The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages indirect injection risks by directing the agent to treat fetched content as source material and never as instructions. Ingestion points for untrusted data occur via web research as defined in SKILL.md. Boundary markers are present in the form of these explicit instructions. The agent's capability inventory includes file operations and command execution for building and serving artifacts. Sanitization relies primarily on these instructional constraints to prevent malicious content from being interpreted as commands.
[DATA_EXFILTRATION]: Instructions in SKILL.md provide a protection against data exposure by requiring that secrets, caches, and private research notes be kept out of the ready-to-serve 'artifact' directory. This ensures that sensitive environment data is not bundled into the portable website.
[SAFE]: The skill contains best-practice security guidance in references/media.md, specifically advising against forwarding credentials to unrelated hosts when using authenticated downloaders for attachments.

oneshot-timeline