The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface.
Ingestion points: The skill reads molecular structure data from files (PDB, SDF, MOL2, mmCIF) provided via command-line arguments in define_binding_site.py and visualize_box.py.
Boundary markers: Absent. The scripts do not use delimiters or instructions to ignore embedded content when parsing external data.
Capability inventory: The skill has the capability to write JSON configuration files and ray-traced image files (PNG). It also executes the PyMOL rendering engine as a subprocess.
Sanitization: Absent. The scripts rely on library-level parsing (MDAnalysis, RDKit, PyMOL) and do not validate the logical integrity of the molecular geometry. An attacker could provide a malicious structure file to influence docking box dimensions or exploit vulnerabilities in the parsing libraries.
[REMOTE_CODE_EXECUTION]: Presence of broken code and unverifiable local dependencies.
Evidence: Both scripts/define_binding_site.py and scripts/visualize_box.py contain code blocks that attempt to import a module named src.utils.config_utils, which is not included in the skill files.
Evidence: The scripts attempt to access several undefined variables, including _params_path, _config, and args.output_dir (which is not defined in the argparse parser).
Detail: This non-functional code appears to be a template artifact or an incomplete port from a larger project. It introduces a security risk because an AI agent might be manipulated into creating a malicious local src directory to resolve the missing dependency, leading to arbitrary code execution when the scripts are run.

drug-binding-site-definition