The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests molecular data from external URLs and cloud storage providers, creating a potential surface for adversarial instructions embedded in file content or metadata to influence agent behavior. \n
Ingestion points: Functions such as dm.read_sdf and dm.open_df documented in references/io_module.md support remote paths. \n
Boundary markers: Documentation in SKILL.md and references/core_workflows.md provides guidelines for confirming remote operations and using cloud paths only when requested. \n
Capability inventory: The skill allows file writing and command execution, supporting remote data persistence. \n
Sanitization: While standardizing chemical structures, the skill does not provide specific sanitization for non-chemical text or instructions in data files. \n- [EXTERNAL_DOWNLOADS]: The skill instructions specify the installation of the datamol package and related cloud backends (s3fs, gcsfs) from official registries. \n- [COMMAND_EXECUTION]: The agent is directed to perform environment setup and package management using uv pip install as described in the SKILL.md installation section. \n- [DATA_EXFILTRATION]: Support for cloud providers (S3, GCS) enables reading and writing data to external endpoints. The documentation mitigates risk by explaining that credentials are used locally and advising path verification.

datamol