The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documentation and API references describe tools for executing shell-level commands. This includes instructions for running system performance benchmarks (NCCL tests) and the ability to execute arbitrary commands on the Modal cloud platform using the cli.cloud.modal_.run_cmd function.
[REMOTE_CODE_EXECUTION]: The framework's documentation includes examples for installing system utilities and ML extensions using standard script-to-shell patterns, such as the uv package manager installer script (curl -LsSf https://astral.sh/uv/install.sh | sh).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and fine-tune models on large-scale datasets from external and untrusted sources like the Hugging Face Hub and various cloud storage providers. This represents a significant attack surface where malicious instructions embedded in dataset segments could influence the agent's actions, particularly given the skill's extensive command execution capabilities.
Ingestion points: Datasets and remote configurations loaded via the datasets or pretraining_dataset YAML fields, as documented in SKILL.md and references/dataset-formats.md.
Boundary markers: While the documentation discusses role-based data masking, it lacks specific instructions or delimiters to prevent the agent from obeying natural language commands found within the raw text of the datasets.
Capability inventory: High-impact capabilities including local shell execution (all_reduce_perf), remote cloud command execution (run_cmd), and extensive network/file system operations during the training loop.
Sanitization: No explicit logic for sanitizing or validating ingested data against prompt injection is identified.
[EXTERNAL_DOWNLOADS]: The skill includes an explicit utility (cli.utils.fetch.fetch_from_github) for downloading and synchronizing code repositories and configuration files from GitHub during execution.
[DYNAMIC_EXECUTION]: The framework utilizes a plugin architecture (integrations.base.BasePlugin) and runtime monkey-patching techniques that dynamically modify and extend the Python execution environment during model loading and training.

axolotl