The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill setup requires the execution of npx flow-nexus@latest, which downloads and runs code from the public npm registry to initialize the MCP server and manage user authentication sessions.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection through its data handling processes.
Ingestion points: Untrusted data enters the agent context through the config parameter in training functions and the input array in prediction functions as defined in SKILL.md.
Boundary markers: There are no delimiters or instructional warnings present to prevent the agent from following malicious instructions potentially embedded in datasets or model configurations.
Capability inventory: The skill allows for complex operations including distributed code execution in E2B sandboxes, network communication with the Flow Nexus API, and model marketplace interactions.
Sanitization: No validation, escaping, or schema enforcement mechanisms are documented for the external data processed during model training or inference.
[DYNAMIC_EXECUTION]: The skill utilizes E2B sandboxes for model training, which involves the dynamic execution of neural network logic. It additionally supports wasmOptimization for performance and daaEnabled for autonomous agent coordination, both of which rely on dynamic code execution environments.

flow-nexus-neural