The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the user to disable a primary security feature in PyTorch 2.6+ by setting the environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 in references/training-evaluation.md. This forces the use of legacy unpickling for model loading, bypassing the default weights_only=True protection. This action creates a significant vulnerability allowing arbitrary code execution if a user is tricked into loading a malicious or tampered model checkpoint.
[COMMAND_EXECUTION]: In references/training-evaluation.md, the skill provides docker run templates using high-privilege flags, including --ipc=host and --ulimit memlock=-1. The --ipc=host flag allows the container to access the host's Inter-Process Communication namespace, which can be leveraged for container escape or data leakage. While ulimits are common for high-performance machine learning, they represent a reduction in host-to-container isolation.
[COMMAND_EXECUTION]: In references/setup-checkpoints.md, the skill executes a multi-line inline Python script using a shell HEREDOC (`python
/data/train.json <<'PY' ...`). This pattern of executing dynamic code blocks is a potential vector for modification if the skill's content were compromised.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface (Category 8) as it processes untrusted transcript data from user-supplied manifests.
Ingestion points: Data enters the system via train.json and val.json manifests described in references/data-lhotse.md.
Boundary markers: There are no documented delimiters or instructions to ignore embedded commands within the transcripts.
Capability inventory: The skill has extensive command execution capabilities through Docker and Python scripts as detailed in references/training-evaluation.md.
Sanitization: There is no evidence of sanitization or filtering of the transcript text before it is processed by the training or evaluation logs, which could allow malicious content to influence agent behavior during error analysis or refinement phases.

nemo-speech-asr-finetune