work-with-kombai
Warn
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: MEDIUMPRIVILEGE_ESCALATIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill provides explicit instructions for the agent to bypass security sandboxes in the Codex environment.
- Instructions explicitly tell the agent to use
sandbox_permissions: "require_escalated"for the first and all subsequent tool calls. - Provides user-facing instructions to enable
danger-full-accessflags to allow the tool to function outside restricted environments. - [EXTERNAL_DOWNLOADS]: The skill fetches and executes external software at runtime via package managers.
- Uses
npx -y @kombai/cli@0.2.0to download and execute the vendor's command-line interface. - [INDIRECT_PROMPT_INJECTION]: The skill context is dynamically expanded using content generated at runtime, creating a surface for instruction injection.
- Ingestion points: Additional skill instructions are ingested via the output of the
kombai skillcommand (found inSKILL.md). - Boundary markers: There are no delimiters or "ignore embedded instructions" warnings surrounding the dynamic output.
- Capability inventory: The skill can execute shell commands (
npx,node,brew,winget), write to the home directory (~/.kombai), and modify environment configuration folders. - Sanitization: No validation or sanitization is performed on the instructions returned by the CLI before the agent follows them.
- [COMMAND_EXECUTION]: The skill executes multiple shell commands to manage environment dependencies and run the core tool.
- Instructions for the agent to install or update Node.js using system-level package managers such as Homebrew (
brew), winget, and nvm. - Direct execution of JavaScript files located in the user's home directory (
$HOME/.kombai/cli/install/latest/launch.mjs). - Execution of
mcp install codexwhich modifies application configuration folders.
Audit Metadata