The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data (text prompts and sampling parameters) which is subsequently processed by an inference microservice.
Ingestion points: User-provided messages and configuration fields in the request body as defined in references/request.yaml.
Capability inventory: The skill executes shell commands (docker, kubectl, ssh), performs network requests (urllib.request), and maintains a local service registry (/tmp/tao-inf-ms-state.json).
Boundary markers: The skill employs standard OpenAI/vLLM chat completion schemas but does not define explicit delimiters to isolate user-provided content within the request structure.
Sanitization: Employs shlex.quote to sanitize user inputs when constructing shell commands for container execution.
[COMMAND_EXECUTION]: Generates and executes shell commands to manage container lifecycles on various platforms using docker run, kubectl, and ssh for remote management.
Commands are constructed using sanitized user inputs such as job_id and network_arch.
Note: Uses umask 0 before container execution, which results in world-writable permissions for any files created by the service within the container's context.
[EXTERNAL_DOWNLOADS]: Fetches container images from the NVIDIA Container Registry (nvcr.io) and model weights from the HuggingFace Hub.
These downloads target well-known and trusted vendor-related services necessary for the skill's functionality.
[DYNAMIC_EXECUTION]: Dynamically generates and executes Python and shell snippets at runtime to facilitate job payload construction, service registry management, and readiness checks.
This mechanism incorporates user-defined parameters such as model paths and architecture choices into executable logic.

tao-run-inference-service