doubao
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill enables the registration and execution of Model Context Protocol (MCP) servers through the
doubao mcp registercommand. The documentation explicitly states that these server processes are not confined by the application's permission sandbox, which allows the agent to execute arbitrary local commands and scripts. - [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, specifically chat messages and turn history from the Doubao service. This creates a potential surface for indirect prompt injection if the ingested content contains malicious instructions.
- Ingestion points: Data retrieved via
doubao sessions read,doubao sessions status, anddoubao sessions currentinSKILL.md. - Boundary markers: No explicit boundary markers or delimiters for ingested message content are defined in the instructions.
- Capability inventory: The skill can execute local commands via MCP registration (
mcp register), manage local files via the workspace flag (--workspace), and perform network operations. - Sanitization: No sanitization or validation logic for the ingested message content is described.
- [CREDENTIALS_UNSAFE]: The skill handles sensitive authentication tokens and cookies used for session management. It supports importing these credentials from local files (
--cookie-file) or environment variables (DOUBAO_HEADLESS_COOKIE). While the instructions caution against printing these secrets, their handling in plain text via CLI arguments or environment variables increases the risk of accidental exposure. - [EXTERNAL_DOWNLOADS]: The instructions direct the user to install the
doubao-clipackage from the npm registry. This package includes an automatic update mechanism (doubao update) which allows for the dynamic modification of the local executable code.
Audit Metadata