The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The knowledge base contains code templates that use high-risk dynamic execution functions to run code generated by the LLM at runtime.
references/knowledge/nlp/aimo-2-2025.md: The execute_python_code function uses exec(code, {}) to run Python strings generated by the model.
references/knowledge/nlp/arc-prize-2025.md: Mentions 'Evolutionary Program Synthesis' where the model fine-tunes itself on search trajectories involving generated code.
[COMMAND_EXECUTION]: Several reference files provide templates for executing system commands through the subprocess module.
references/knowledge/nlp/aimo-2-2025.md: The execute_code_safely function uses subprocess.run(['python', temp_file], ...) to execute a temporary file populated with model-generated code.
references/knowledge/nlp/konwinski-prize-2025-comparison.md: The _run_test function uses subprocess.run(['pytest', test_file, '-v'], ...) on dynamically created test files.
[DATA_EXFILTRATION]: The skill provides patterns for performing network operations to external domains to collect data.
references/knowledge/nlp/aimo-2-2025.md: The build_openmath_dataset function uses the requests library to scrape https://artofproblemsolving.com.
[PROMPT_INJECTION]: The skill is designed to ingest untrusted data from Kaggle competition URLs and apply the extracted techniques to its own operations, representing an indirect prompt injection surface.
Ingestion points: Untrusted Kaggle competition data processed via URL (specified in SKILL.md).
Boundary markers: Absent; no instructions are provided to the agent to ignore or delimit instructions found within the competition data.
Capability inventory: Includes exec, subprocess.run, and requests calls across multiple reference files (e.g., aimo-2-2025.md, konwinski-prize-2025-comparison.md).
Sanitization: Absent; the code templates do not implement sanitization for external content before interpolation into prompts or execution environments.

kaggle-learner