agent-workbench
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill dynamically fetches its core operating rules and instructions from a remote MCP server, creating a surface for indirect prompt injection.
- Ingestion points:
SKILL.mdinstructs the agent to search for and read canonical instructions from MCP resources atagent-protocol/70-agent-workbench.mdandai-workspace/60-skills/agent-workbench/SKILL.md. - Boundary markers: Absent. The instructions lack delimiters or warnings to ignore potentially malicious embedded instructions in the remote content.
- Capability inventory: The agent utilizes MCP tools like
resource_read_budgeted,context_prepare, andresource_get_contentto fetch and act upon remote data. - Sanitization: Absent. The skill does not describe any validation or filtering mechanisms for the dynamically loaded instructions.
- [COMMAND_EXECUTION]: The
workbench.mjsmodule executes the localcodexCLI tool usingnode:child_process.spawnSyncto manage and validate MCP server configurations. - [EXTERNAL_DOWNLOADS]: The
doctorfunctionality inworkbench.mjsperforms network operations to a remote MCP server (defaulting to the author's domainagent-mcp.devkimson.kro.kr) to verify tool availability and authentication status.
Audit Metadata