The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: Several sub-skills are vulnerable to indirect prompt injection as they ingest untrusted external data to be processed by LLMs.
Evidence: The chemeagle-guide processes external chemistry papers (PDF or text format) using a multi-agent pipeline (Document Parser, Chemical NER, Reaction Extraction) to build structured databases. Maliciously crafted papers could attempt to override agent instructions.
Evidence: The cactus-cheminformatics-guide and chemgraph-agent-guide provide natural language interfaces (agent.ask, agent.run) for molecular analysis and simulation workflows. If these agents are used to process user-supplied queries or data without strict input sanitization, they could be manipulated.
Mitigation: chemeagle-guide includes a 'Validation Agent' to cross-check extracted data, which partially mitigates the risk, though it does not eliminate the prompt injection surface.
[EXTERNAL_DOWNLOADS]: The guides instruct users to download and install various third-party scientific packages and containers.
Evidence: chemgraph-agent-guide/SKILL.md suggests pip install chemgraph and docker pull ghcr.io/argonne-lcf/chemgraph:latest.
Evidence: Various guides reference packages such as rdkit, nmrglue, pyteomics, MDAnalysis, and openmm.
Context: These are standard, well-known scientific libraries in the chemistry domain, but they represent a dependency chain that users must trust.
[COMMAND_EXECUTION]: The skill provides templates for executing specialized chemistry CLI tools.
Evidence: molecular-dynamics-guide/SKILL.md provides shell commands for GROMACS (gmx pdb2gmx, gmx mdrun, etc.).
Evidence: computational-chemistry-guide/SKILL.md provides commands for running ORCA calculations.
Context: This is the intended primary purpose of the skill (automating research workflows) and is acceptable within a controlled research environment.

chemistry-skills