The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates and executes Python code (experiment.py) at runtime using local subprocesses (sandbox mode) or remote SSH connections (ssh_remote mode). This pattern of generating and then executing scripts represents a significant attack surface.
[INDIRECT_PROMPT_INJECTION]: The skill orchestrates an autonomous pipeline where external research data (literature review) can influence code generation, creating a vulnerability surface where malicious instructions in research papers could lead to the generation of harmful code.
Ingestion points: User-provided research topic and external literature data fetched during the pipeline execution (SKILL.md).
Boundary markers: The instructions do not specify any delimiters or warnings to ignore embedded instructions in the ingested data.
Capability inventory: The skill utilizes file system access, shell command execution via the researchclaw CLI, and subprocess spawning for experiment execution.
Sanitization: No sanitization or validation of the generated code is described before execution.
[COMMAND_EXECUTION]: The skill requires the execution of the researchclaw CLI tool and various bash commands to manage the research workflow, including file management and pipeline orchestration.
[AUTONOMY_CONCERN]: The --auto-approve option enables the agent to bypass manual review gates at stages 5, 9, and 20, which are intended to provide human oversight for the generated research plan and experiment code before they are acted upon.

researchclaw