The Agent Skills Directory

[DYNAMIC_EXECUTION]: The eval/run-eval.sh script contains a vulnerability in the eval_assertion function where values extracted from the AI agent's responses are interpolated directly into code strings executed via python3 -c and node -e. This lack of sanitization allows a manipulated AI response to execute arbitrary shell commands on the host system running the evaluation suite.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a natural language interface to generate structured configuration files, creating a vulnerability to indirect prompt injection.
Ingestion points: User-supplied natural language descriptions for ComfyUI workflows (SKILL.md).
Boundary markers: None present; the skill does not use delimiters or include instructions for the agent to ignore directives within the input data.
Capability inventory: The skill can write files to the local filesystem (in the projects/ directory) and call the comfyui-api tool.
Sanitization: None; there is no validation or filtering of the input prompt before it is used to generate executable node graphs.
[COMMAND_EXECUTION]: The evaluation framework in eval/run-eval.sh executes the claude CLI and several utility binaries (sed, grep, python3, node). The unsafe handling of agent output within these execution contexts represents a significant command injection risk for the user's local machine.

comfyui-workflow-builder