The Agent Skills Directory

[COMMAND_EXECUTION]: The research_experiment_update_execution tool permits the agent to define arbitrary shell command lists and environment variables. These commands are subsequently executed by the research_experiment_launch tool, creating a vector for arbitrary code execution if the parameters are manipulated. Other tools like research_workflow_task_execute and research_project_closure_execute also trigger automated execution routines.
[DATA_EXFILTRATION]: Tools such as research_artifact_upsert and research_dataset_version_create take file system paths as parameters, which could allow an agent to be tricked into reading and exposing sensitive local files by incorporating them into the research graph.
[PROMPT_INJECTION]: The skill retrieves unstructured data from research notes and memories, presenting an indirect prompt injection surface.
Ingestion points: Content is retrieved using research_notes_search and research_project_memory_list in __init__.py.
Boundary markers: The skill does not implement delimiters or safety instructions when processing retrieved research data.
Capability inventory: The skill possesses significant capabilities for file system interaction and shell command execution.
Sanitization: No input validation or sanitization is performed on the data retrieved from the research graph before it is returned to the agent's context.

research-workflows