The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/align_encoder.py invokes several HuggingFace model loading methods (e.g., AutoTokenizer.from_pretrained, AutoModel.from_pretrained, and AutoFeatureExtractor.from_pretrained) with the trust_remote_code=True parameter enabled. This configuration allows for the execution of arbitrary code defined within a model's repository. While common for certain model architectures, it creates a potential vulnerability if the agent is instructed to benchmark untrusted or malicious model IDs.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of several external Python dependencies, including tribev2, transformers, torch, scikit-learn, numpy, opencv-python-headless, and soundfile. Furthermore, the workflow involves downloading large model weights (approximately 10 GB) from HuggingFace repositories, specifically facebook/tribev2.
[COMMAND_EXECUTION]: The SKILL.md file and integration documentation provide multiple shell command examples for the agent to execute, including dependency installation via pip and the execution of the Python benchmarking script. This functional pattern assumes the agent has shell access and relies on the integrity of the provided command-line arguments.
[DATA_EXFILTRATION]: The script scripts/align_encoder.py is designed to read files from a user-specified stimulus directory (--stimulus-dir). While intended for processing text, video, and audio data for alignment scoring, this functionality involves accessing and reading local filesystem content, which could be targeted for unauthorized data access if stimulus paths are not properly constrained.

tribe-v2-agent-alignment