The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs users to install its management tool by piping a remote script directly to a shell (curl -fsSL https://sup.soma.org | bash). Although this script is hosted on the vendor's own domain, shell-piping remote content is an unverified execution pattern that bypasses standard package manager safety checks.
[EXTERNAL_DOWNLOADS]: The workflow requires downloading several specialized libraries and datasets, including the soma-sdk and soma-models packages, and streaming the 'The Stack v2' dataset from HuggingFace for submission scoring.
[COMMAND_EXECUTION]: The skill relies on extensive use of the soma CLI and the modal cloud platform to orchestrate GPU-intensive training and on-chain signing operations.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface because it ingests and processes untrusted source code and text from external datasets (e.g., The Stack v2) for next-byte prediction tasks.
Ingestion points: Untrusted source code is streamed into the agent's context via the stream_stack_v2 function described in the submitter workflow.
Boundary markers: No explicit delimiters or 'ignore' instructions are provided to the agent to separate dataset content from instruction logic during processing.
Capability inventory: The skill can execute shell commands via the soma CLI, interact with the Modal cloud environment, and write data to public S3 buckets.
Sanitization: External content is processed as raw bytes for scoring and training without filtering for potential malicious natural language instructions.

soma