concept-implementation
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill identifies a surface for indirect prompt injection because it is designed to ingest and act upon untrusted external concept models and application specifications during its implementation phase.\n
- Ingestion points: As defined in
SKILL.md(Step 1), the agent readsreferences/spec-format.md, application purpose documents, and concept models to begin the implementation process.\n - Boundary markers: The skill does not provide instructions for the agent to use specific delimiters or to treat the ingested model content as untrusted data, which could allow instructions embedded in those models to override the agent's behavior.\n
- Capability inventory: The agent is granted capabilities to generate implementation code across multiple languages (Rust, Java, TypeScript), define architectural plans, and modify the repository file structure.\n
- Sanitization: There are no explicit requirements or steps for the agent to sanitize, validate, or filter the content of the external models and specifications before they are used for code generation and decision-making.
Audit Metadata