The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines workflows for training Sparse Autoencoders (SAEs) that involve processing large external datasets, creating a surface for potential indirect prompt injection if the source data contains malicious instructions.
Ingestion points: SKILL.md (Workflow 2) and references/api.md describe training configurations using dataset_path="monology/pile-uncopyrighted" to ingest external content.
Boundary markers: The provided implementation templates do not use delimiters or explicit instructions to prevent the model from obeying instructions that might be embedded within the training corpus.
Capability inventory: The skill utilizes network operations via SAE.from_pretrained (to fetch model artifacts from HuggingFace) and performing file system writes using sae.save_model to store trained weights.
Sanitization: There is no evidence of sanitization or content validation for the data ingested during the training process, which is typical for library-focused research code but remains a risk factor.

sparse-autoencoder-training