ai-features
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill probes local network services, specifically Ollama on port 11434 and LM Studio on port 1234, to detect available AI runtimes on the user's machine during the inventory phase.
- [CREDENTIALS_UNSAFE]: The skill identifies and checks for the existence of API keys (Anthropic, OpenAI) in environment variables and configuration stores. It mandates security best practices, such as truncating keys when displayed in UI and ensuring they are never committed to the repository files.
- [EXTERNAL_DOWNLOADS]: The skill references well-known packages, model registries (Ollama library), and official documentation from trusted sources (Anthropic, OpenAI, Hugging Face) for downloading AI models and dependencies.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests application source code, schemas, and migrations to propose AI features. It proactively addresses injection risks by instructing the agent to place untrusted content in structured data blocks (JSON) and avoid including it in system prompts, citing OWASP LLM Top 10 guidelines.
- [DYNAMIC_EXECUTION]: The skill describes using SQLite extensions (sqlite-vec) and loading AI models at runtime via Transformers.js or external daemons, following standard and documented implementation patterns for AI integration.
Audit Metadata