The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: In references/architecture.md, the code demonstrates fetching and executing code from the kernels-community/flash-attn3 repository using a get_kernel function, which constitutes unverified remote code execution from a source outside the verified scope.
[EXTERNAL_DOWNLOADS]: The references/experiment-loop.md file references an import of rustbpe, a non-standard and unverifiable Python package that is not available in primary public registries.
[COMMAND_EXECUTION]: The autonomous experiment workflow in references/experiment-loop.md instructs the agent to modify source code (train.py) and execute shell commands including git commit, git reset, and training script invocation, creating a significant capability surface for exploitation.
[PROMPT_INJECTION]: The skill is designed to process data from external sources (train_loader, data_source) while having file-writing and shell-execution capabilities.
Ingestion points: Data ingestion via train_loader in SKILL.md and data_source in references/experiment-loop.md.
Boundary markers: None identified; no instructions exist to isolate processed data from the agent's control flow.
Capability inventory: The skill can write scripts, execute shell commands, and terminate processes via exit() calls.
Sanitization: No validation or sanitization of ingested data or the resulting code modifications is present.

ml-training-recipes