The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes numerous shell commands and scripts (e.g., conda-envs/*/install.sh, install_lammps.sh) to build the LAMMPS environment and its MLIP backends. These scripts are referenced but not provided in the analyzed file set.
[EXTERNAL_DOWNLOADS]: examples/mace/run_mace_na2si3o7_quench.sh fetches MACE model checkpoints from Hugging Face (a well-known service) via the download_mace_mp_checkpoint function and direct URLs.
[REMOTE_CODE_EXECUTION]: examples/matgl/run_matgl_cu_phase_transition.sh executes pip install at runtime to install numpy<2 and chgnet, which involves downloading and installing external code from public registries.
[DYNAMIC_EXECUTION]: examples/mace/run_mace_na2si3o7_quench.sh uses torch.load to deserialize model files and torch.jit.compile to generate executable code for the MACE backend. torch.load is known to be vulnerable to arbitrary code execution if the input file is malicious.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect injection (Category 8) by processing user-provided structure files (INPUT_STRUCTURE) and model names.
Ingestion points: examples/mace/run_mace_na2si3o7_quench.sh reads external files via ase.io.read.
Boundary markers: None present in the prompt templates.
Capability inventory: Access to lmp binary execution and file system writes in OUT_DIR across multiple scripts.
Sanitization: No validation or sanitization of external structure file content before processing.

mat-lammps-md