product-help-center
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill represents a vulnerability surface as it is designed to ingest and process data from external sources (defined in
data/sources.json) and user-provided requirements to generate AI-driven documentation systems. - Ingestion points: Data enters the agent's context through the
sources.jsoncatalog and user-supplied design requirements inSKILL.md. - Boundary markers: The skill explicitly mandates a 'Retrieval-first' policy in
references/ai-integration.md, requiring citations for factual claims and stating that assistants should answer only from retrieved evidence. - Capability inventory: The skill utilizes local scripts for operations, including
append_learning.pyfor persistence of learnings andvalidate_sources.pyfor network-based validation of sources. - Sanitization: No explicit sanitization logic is provided for the internal feedback scripts, though they are scoped to local file operations.
- [EXTERNAL_DOWNLOADS]: The
data/sources.jsonfile contains a curated list of external URLs for official documentation and protocol specifications. - Evidence: The skill references domains such as
support.zendesk.com,intercom.com,docs.readme.com,gitbook.com,mintlify.com,developers.google.com,learn.microsoft.com, anddocs.stripe.com. - Assessment: All referenced external sources are official documentation portals for well-known and established technology services and cloud providers, which are treated as safe under security protocols.
- [COMMAND_EXECUTION]: The skill includes instructions and scripts for performing maintenance and operational tasks.
- Evidence:
SKILL.mdinstructs the agent to useagents-skills-feedback-loop/scripts/append_learning.pyto record new patterns, andscripts/validate_sources.pyprovides a utility to validate the integrity of the source catalog using Python's standardurlliblibrary. - Assessment: These are standard operational tools for maintaining skill state and data integrity, executing locally within the agent's environment without processing untrusted shell input.
Audit Metadata