The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from the target environment that could contain malicious instructions designed to subvert the agent's behavior.
Ingestion points: The skill reads the ADMIN_MCP_INSTRUCTIONS field via the site_setting-get tool, as described in SKILL.md and references/checklist.md (Item B11).
Boundary markers: The instructions include a specific rule to quote the text and verify its intent with the user, acting as a manual boundary to identify unexpected instructions.
Capability inventory: The skill is strictly limited to read-only MCP tools for site settings, member lists, and logs, preventing high-impact write operations from being triggered by injected content.
Sanitization: No automated sanitization of the retrieved metadata is performed before the agent processes it, relying instead on the manual human-in-the-loop verification step.

kuroco-security-audit