concept-synthesis
Pass
Audited by Gen Agent Trust Hub on Aug 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes raw concept stubs that potentially contain data from external sources (e.g., web articles ingested via signal-detector). This data is fed into an LLM to generate summaries, tiers, and curation verdicts, creating a surface for indirect prompt injection.
- Ingestion points: Concept pages located in the
concepts/directory, which originate from ingestion pipelines (SKILL.md). - Boundary markers: The instructions do not specify explicit delimiters or "ignore embedded instructions" warnings for the LLM synthesis phase.
- Capability inventory: The skill possesses significant capabilities including reading thousands of pages, writing new synthesized pages (
gbrain put), and deleting pages (gbrain delete) (SKILL.md). - Sanitization: The skill includes a "Grounding labels" mechanism (Phase 5) to flag opinionated or unsafe claims, which acts as a rudimentary validation layer but does not prevent the underlying risk of instructions hidden in ingested text being followed by the agent.
Audit Metadata