The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources into agent prompts without robust boundary markers or sanitization. * Ingestion points: Data enters the context via variables such as survey_res (user ideas), references (external papers), code_survey_res (survey notes), and judge_res (feedback loop) as seen in SKILL.md and the templates in the prompts/ directory. * Boundary markers: The prompt templates use simple text labels but lack strong delimiters or instructions to ignore embedded commands. * Capability inventory: The ML Agent (documented in references/ml_agent_instructions.md) has access to tools including execute_command (shell access), run_python, write_file, and package managers (pip, conda). * Sanitization: There is no evidence of input validation or escaping for the interpolated variables.
[EXTERNAL_DOWNLOADS]: The Machine Learning Agent is explicitly instructed to install missing dependencies using pip install or conda install during the implementation phase, which involves downloading unverified third-party packages.
[COMMAND_EXECUTION]: The workflow involves the execution of shell commands and Python scripts that are dynamically generated or modified by the agent itself.
[DYNAMIC_EXECUTION]: The skill requires the agent to write project structures and then execute the resulting code. In the submission phase, the agent is instructed to dynamically modify existing scripts to adjust parameters and add checkpoint-saving logic before execution.

inno-experiment-dev