The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches pretrained model weights from HuggingFace, which is a well-known and trusted service. The code uses TribeModel.from_pretrained("facebook/tribev2") to download these resources.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided media files (video, audio, text) using Large Language Models (LLaMA 3.2). This creates an ingestion surface where adversarial content embedded in the stimuli could potentially influence the model's behavior.
Ingestion points: Data enters the system via video_path, audio_path, and text_path parameters in the get_events_dataframe method as documented in SKILL.md.
Boundary markers: No explicit delimiters or instructions are present in the skill to treat the content of these external files as untrusted or to ignore embedded instructions.
Capability inventory: The skill facilitates brain response predictions and model training; it does not demonstrate capabilities for arbitrary system command execution or network exfiltration of sensitive user data.
Sanitization: There is no documentation of sanitization or filtering of the content within the input media files before they are processed by the multimodal encoders.

tribev2-brain-encoding