designing-agent-systems
Pass
Audited by Gen Agent Trust Hub on Sep 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided design goals, constraints, and architecture shapes as untrusted data, which presents a surface for indirect prompt injection.\n
- Ingestion points: User-provided JSON objects containing design problems are interpolated into the
pattern-librariandispatch prompt inSKILL.md.\n - Boundary markers: The skill implements robust defensive measures, using
BEGIN_UNTRUSTED_DESIGN_PROBLEMandEND_UNTRUSTED_DESIGN_PROBLEMmarkers and explicit instructions to the agent to treat the content as inert data.\n - Capability inventory: The skill has the capability to write files (scaffolding artifacts) and invoke MCP tools based on the results of the consultation.\n
- Sanitization: The instructions strictly prohibit following any embedded instructions or links found within the untrusted payload.\n- [COMMAND_EXECUTION]: The skill executes Model Context Protocol (MCP) operations (e.g.,
atlas_orient,atlas_cards,atlas_read) to interact with the external pattern corpus. It also performs filesystem operations to scaffold new agent systems upon user approval.\n- [EXTERNAL_DOWNLOADS]: The skill references the Agentic Atlas documentation site (agentic-atlas.dev) for citations. These links are provided for user reference and are not fetched as executable content or control instructions.
Audit Metadata