The Agent Skills Directory

[COMMAND_EXECUTION]: The agent is designed to generate and run code with system privileges to perform complex biomedical tasks, which could lead to unauthorized system access if the generated code is malicious or flawed.
[REMOTE_CODE_EXECUTION]: The framework utilizes LLM-driven reasoning to dynamically create and execute code pipelines at runtime, a pattern that inherently allows for arbitrary code execution.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of a significant 11GB data lake of biomedical knowledge from Stanford SNAP lab resources and installs the biomni package from public registries.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it reads and processes external files like single-cell RNA-seq datasets and GWAS summary statistics which could contain adversarial instructions.
Ingestion points: Processes user-provided file paths for genomic and clinical datasets.
Boundary markers: No specific boundary markers or instruction isolation techniques are documented for data ingestion.
Capability inventory: Possesses full code execution and file system manipulation capabilities.
Sanitization: Relies on manual code review and environment isolation rather than automated sanitization of ingested content.

biomni