The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes torch.load() and solver.load() to manage model checkpoints (e.g., solver.pth, pretrained.pth). In TorchDrug 0.2.1, these functions rely on PyTorch's default pickle-based deserialization, which can be used to execute arbitrary code if the checkpoint file is sourced from an untrusted origin or tampered with locally.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external scientific data formats which could contain adversarial instructions.
Ingestion points: The skill reads external data through datasets.ClinTox, Protein.from_pdb, and Molecule.from_smiles across SKILL.md and references/core_concepts.md.
Boundary markers: The instructions lack explicit delimiters or "ignore embedded instructions" warnings when handling scientific strings or file contents.
Capability inventory: The agent has access to Bash, Read, Write, and Edit tools, allowing for high-impact actions if an injection occurs.
Sanitization: There is no evidence of string sanitization or schema validation for external data before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to download binary wheels from https://data.pyg.org/whl/torch-2.0.0+cpu.html. This targets the official PyTorch Geometric wheel registry, which is a well-known service in the graph-learning ecosystem.

torchdrug