The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill directs the agent to "Create temporal python files" and "Execute this python code" in the root folder to validate assumptions and test logic. This script generation and execution happens at runtime and is driven by the user's request. While intended for research, it constitutes a high-privilege execution pattern.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon untrusted user requests without explicit boundary markers or sanitization, potentially allowing malicious instructions to influence the agent's actions.
Ingestion points: The user request is the primary input source for the research task.
Boundary markers: No markers or instructions are provided to the agent to treat user input as untrusted or to ignore embedded instructions during the code creation phase.
Capability inventory: The skill instructions allow for checking files, creating new Python scripts, executing code, and deleting files.
Sanitization: No sanitization, filtering, or validation logic is defined for the input or the resulting generated code before it is executed.
[COMMAND_EXECUTION]: The instructions explicitly forbid user interaction ("DO NOT ask for clarifications or wait for input"), which removes the human-in-the-loop oversight usually required for executing generated code, increasing the risk of the agent executing harmful operations autonomously if influenced by a malicious request.

research