The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The documentation provides examples and instructions for the agent to use sensitive credentials such as API keys, Access Key IDs, and Access Key Secrets via command-line flags (e.g., --api-key, --access-key-id, --access-key-secret). While this is standard for the CLI tool, passing secrets as command-line arguments can expose them in process logs or history.
[REMOTE_CODE_EXECUTION]: The skill includes instructions for installing, initializing, and updating external code via the bl CLI. Commands such as bl skill add, bl skill init, and bl plugin install allow the agent to download and integrate new skills or command packs from the official Bailian registry.
[COMMAND_EXECUTION]: The primary function of this skill is to execute arbitrary bl commands on the host system to manage cloud resources and perform tasks such as model inference and web searches.
[DYNAMIC_EXECUTION]: The bl pipeline run command enables the execution of workflow definitions from local files (YAML/JSON). This allows for the dynamic execution of logic defined outside the static skill instructions.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with external and potentially untrusted data through search results, knowledge base retrieval (RAG), and MCP tool outputs.
Ingestion points: Untrusted data enters the context through bl search web, bl knowledge search/chat, and bl mcp call commands.
Boundary markers: Not explicitly specified in the documentation files; the agent relies on its internal safety protocols.
Capability inventory: The bl CLI has capabilities including network communication, file system operations (upload/read), and execution of external scripts (skill installation).
Sanitization: Sanitization of external content before interpolation is not detailed in the command reference.

bailian-cli