The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses importlib.import_module to load Python modules based on strings extracted from external config.json files, creating a risk if the configuration file is compromised.
[REMOTE_CODE_EXECUTION]: Instructions explicitly guide the agent to load models using the trust_remote_code=True flag, which executes arbitrary Python code provided by the model author.
[CREDENTIALS_UNSAFE]: The skill uses and references sensitive credentials, including SSH private keys (~/.ssh/id_rsa) for remote cluster access and HF_TOKEN for accessing gated datasets.
[COMMAND_EXECUTION]: The skill performs shell command execution locally and on remote SLURM clusters, including job submission and environment inspection using a custom Python launcher.
[EXTERNAL_DOWNLOADS]: The skill downloads container images from NVIDIA's registry (nvcr.io), clones source code from Hugging Face's GitHub repositories, and installs various third-party Python packages from PyPI.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted external model metadata and source code (modeling_*.py), which serves as an attack surface for influencing the agent's behavior via data-driven instructions.

ptq