The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes structured data from the TensorArt API that could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: Data enters the context through scripts/list_tools.py (which fetches a tool list) and scripts/query_task.py (which fetches task results and status).
Boundary markers: Absent; there are no specific instructions or delimiters to isolate the API response content from the agent's command logic.
Capability inventory: The agent has access to Bash (running local scripts), Read, and Write capabilities.
Sanitization: None; the scripts return JSON data directly to the agent without filtering for potential prompt injection patterns.
[EXTERNAL_DOWNLOADS]: The skill includes a utility scripts/download_result.py designed to download files from remote URLs to local paths.
The script uses urllib.request to fetch content from any provided URL and write it to a specified destination (defaulting to /tmp).
While intended for downloading generation results from the vendor's infrastructure, the script does not implement domain validation or URL whitelisting, which could be exploited if an attacker can manipulate the URL provided to the agent.
[COMMAND_EXECUTION]: The skill relies on executing local Python scripts (list_tools.py, create_task.py, etc.) to perform its core functions.
The execution is restricted to the skill's own directory via Bash(python3 *), following the principle of least privilege, but this represents the primary capability set that could be abused via indirect injection.
[SAFE]: The management of the ~/.tensor_access_key is handled as a standard configuration practice for CLI-based skills. The instructions guide the user to store their own key in a dedicated hidden file, which is a recognized method for secret management in this context.

tensorart-generate