The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the creation of environments where agents process untrusted data, such as task prompts, system instructions, and live model-driven user simulations, which creates a surface for indirect prompt injection.
Ingestion points: External data is ingested via vf.TaskData (prompts, system prompts, reference answers) and through the vf.Env interaction loop where agents respond to inputs from other agents or simulated users.
Capability inventory: The framework allows for subprocess execution (runtime.run_program, runtime.run_uv_script), custom MCP tool registration, and containerized task execution.
Boundary markers: The skill notes the use of an InterceptionServer to capture traces and model requests, providing a level of visibility, but does not provide specific instructions for delimiting untrusted inputs.
Sanitization: The provided templates do not include explicit input sanitization or validation logic for task-specific prompts before they are processed by the agents.

create-environments