The Agent Skills Directory

[COMMAND_EXECUTION]: The technical analysis for the 'AIMO-2' competition in references/knowledge/nlp/aimo-2-2025.md contains code implementations for the 'MARIO' framework that use subprocess.run to execute Python scripts generated from untrusted LLM outputs.
[REMOTE_CODE_EXECUTION]: Multiple files, such as references/knowledge/nlp/aimo-2-2025.md, contain code templates employing the exec() function to run dynamically generated Python strings. These strings are constructed from data retrieved via network requests to external forums (e.g., AoPS), creating a path for arbitrary code execution.
[PROMPT_INJECTION]: The skill's core 'self-evolving' mechanism creates a broad surface for indirect prompt injection.
Ingestion points: External data enters the agent context through user-provided Kaggle URLs and automated scraping of third-party forums mentioned in the knowledge base (e.g., aimo-2-2025.md).
Boundary markers: The knowledge files lack delimiters or protective instructions to prevent the agent from accidentally or maliciously obeying commands embedded in ingested data.
Capability inventory: The skill possesses powerful capabilities including filesystem persistence (writing to the references/ directory), network scraping via requests, and dynamic code execution via subprocess and exec.
Sanitization: No input sanitization or validation logic is implemented in the provided code patterns to filter external content before processing.
[DATA_EXFILTRATION]: The templates in references/knowledge/nlp/aimo-2-2025.md and references/knowledge/time-series/birdclef-plus-2025.md demonstrate the use of the requests library to fetch content from external domains. When combined with templates that utilize the openai API, these capabilities could be misused to transmit harvested local data or sensitive environment variables to external servers.

kaggle-learner