The Agent Skills Directory

[COMMAND_EXECUTION]: The skill grants execution permissions to and runs a local shell script (chmod +x build.sh && bash build.sh). This allows for arbitrary command execution within the agent's shell environment.
[COMMAND_EXECUTION]: It performs installation of locally generated Python wheels using pip install --force-reinstall --no-deps. Installation of package files can trigger the execution of setup scripts or hooks.
[COMMAND_EXECUTION]: The skill dynamically generates Python test files (tests/test_<op_name>.py) and subsequently executes them using python and pytest.
[REMOTE_CODE_EXECUTION]: It uses torch.ops.load_library() to dynamically load compiled shared libraries (.so files) from the project directory. Loading binary libraries from non-static paths is a potential vector for arbitrary code execution if the file path is manipulated.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection.
Ingestion points: The skill ingests CANN installation paths, Conda environment names, and error logs from compilation outputs.
Boundary markers: No specific delimiters or instructions are used to separate untrusted data from execution logic.
Capability inventory: The skill possesses extensive capabilities including file writes, shell command execution (bash, chmod), and package installation (pip).
Sanitization: There is no evidence of sanitization for the environment paths or generated code logic before execution.

external-gitcode-ascend-ascendc-operator-compile-debug