The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill includes a Python helper function bake_and_replace_driver that uses exec() to dynamically set properties on Blender objects. This pattern is often used for automation but carries risk if the input string is not strictly controlled.
Evidence in SKILL.md: exec(f"target.{data_path} = {values[i]}") in the bake_and_replace_driver function.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using curl to send Python scripts to a local Blender process.
Evidence in SKILL.md: curl -s localhost:5656 --data-binary @- <<'PYEOF' used as the primary communication method.
[INDIRECT_PROMPT_INJECTION]: The skill contains a vulnerability surface where user-controlled input (such as property names or data paths) could reach a dynamic execution context (exec) if the agent follows the provided helper templates without adding its own validation logic.
Ingestion points: User-provided object names, property paths, or animation data processed by the agent in SKILL.md.
Boundary markers: The skill does not define specific delimiters or instructions for the agent to treat user-provided names as untrusted data.
Capability inventory: The agent has the capability to execute arbitrary Python code within the Blender environment (via the curl to localhost mechanism), which has access to system modules like os and subprocess.
Sanitization: The provided Python example does not include sanitization or validation for the data_path variable before it is interpolated into the exec() call.

blender-3d