The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill clones a software repository from an external GitHub account (dauparas/LigandMPNN) that is not part of the established trusted organizations list.
[REMOTE_CODE_EXECUTION]: The skill proceeds to execute a shell script (get_model_params.sh) and a Python script (run.py) directly from the newly cloned remote repository, creating a significant dependency on the security of that external source.
[COMMAND_EXECUTION]: The skill uses shell commands like 'sed' to modify downloaded files and 'bash' to run setup scripts, which are executed in the local environment.
[DYNAMIC_EXECUTION]: The documentation notes that one dependency (ProDy) may require runtime compilation from source using 'gcc' or 'g++' if compatible binaries are not found, which involves invoking a compiler during the installation process.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external files (PDB and JSON) which represent an attack surface for indirect prompt injection.
Ingestion points: Reads protein/ligand structures via --pdb_path and amino acid constraints from JSON files.
Boundary markers: No explicit markers or instructions are provided to the agent to isolate or ignore potentially malicious content within these data files.
Capability inventory: The environment allows for package installation, shell script execution, and file modification.
Sanitization: The skill relies on the third-party model code for parsing without additional validation layers documented in the instructions.

ligandmpnn