ollama-delegate
Pass
Audited by Gen Agent Trust Hub on Jul 5, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The script "scripts/ollama-chat.sh" executes a Python wrapper to communicate with the local Ollama HTTP API. Parameters such as model names and file paths are handled securely using shell quoting and Python's "sys.argv" interface to prevent injection.
- [PROMPT_INJECTION]: The skill manages an indirect prompt injection surface by ingesting and processing data generated by an external local model.
- Ingestion points: The local model's response is captured and written to a file ("OUT_FILE") specified by the user.
- Boundary markers: The instructions specify using prompt-based constraints (e.g., "Output ONLY the code") and recommend post-processing to remove markdown fences.
- Capability inventory: The script performs local file writes and interacts with the network at "localhost:11434". The accompanying instructions direct the agent to execute the resulting code for testing and verification purposes.
- Sanitization: The skill relies on the agent's internal reasoning and manual review processes rather than programmatic sanitization to validate the untrusted output from the local model.
Audit Metadata