The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill contains an indirect prompt injection attack surface due to untrusted data ingestion instructions.
Ingestion points: In SKILL.md, the agent is instructed to connect to an external network resource (https://arxiv.org/abs/2609.00065 or http://export.arxiv.org/api/query?id_list=2609.00065) to parse citation authors, year, and version details at runtime.
Boundary markers: No explicit delimiters or boundary isolation techniques are used to separate the external webpage text from the agent's execution context.
Capability inventory: The skill is configured with broad tool capabilities including Bash, Read, Write, and Edit tools.
Sanitization: There is no validation or sanitization process defined to filter potential natural language instructions embedded within the fetched external record.
[EXTERNAL_DOWNLOADS]: The skill workflows in scripts/transfer_learning.py and references/workflows.md configure the automated retrieval of pretrained transformer models (such as seyonec/ChemBERTa-zinc-base-v1 and ibm/MoLFormer-XL-both-10pct) from Hugging Face's official model hub repository via the DeepChem library integration.

deepchem