The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted biological sequence data from FASTA files for processing, creating a surface for potential indirect prompt injection.
Ingestion points: FASTA files (complex.fasta) containing sequence strings are passed to the run_inference function.
Boundary markers: The skill does not employ specific delimiters or instructions to mitigate instructions embedded within sequences.
Capability inventory: The skill utilizes the chai-lab library, requires GPU access, and writes prediction outputs to the filesystem.
Sanitization: No explicit input validation or sanitization of sequence data is detailed in the documentation.
[EXTERNAL_DOWNLOADS]: The tool downloads approximately 5 GB of model weights and language model embeddings from the vendor's GitHub repository upon first use.
[DATA_EXFILTRATION]: Biological sequences are sent to the public ColabFold MSA server (api.colabfold.com) when the MSA option is enabled, which is a standard procedure for this modeling workflow.

chai1