cwcli

Warn

Audited by Gen Agent Trust Hub on Sep 21, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONPERSISTENCECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute the 'caffeinated-whale-cli' package directly from the PyPI registry using the 'uvx' utility. This results in the execution of third-party code in the local environment that does not originate from a recognized trusted vendor.
  • [PERSISTENCE]: The 'setup' verb installs a 'SessionStart' hook into AI agent harnesses such as Claude Code, Codex, and OpenCode. This modifies the agent's initialization process to automatically execute context-gathering commands in every subsequent session.
  • [COMMAND_EXECUTION]: The skill facilitates the execution of shell commands through 'uvx' and 'cwcli' to perform lifecycle management on Docker containers, run database migrations, and manage local application benches.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes data from local development environments which could serve as an attack surface for indirect instructions.
  • Ingestion points: Docker container metadata, status reports, and bench application lists are ingested into the agent context via 'SKILL.md'.
  • Boundary markers: No specific delimiters or safety instructions are defined to separate tool output from instructions.
  • Capability inventory: Provides the ability to start, stop, and remove Docker containers and sites, as well as execute tests and migrations.
  • Sanitization: The skill relies on the 'cwcli' tool's TOON format for structured data but lacks additional sanitization to prevent the agent from interpreting output as instructions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 21, 2026, 02:43 AM
Security Audit — agent-trust-hub — cwcli