Agent Design 2

Pass

Audited by Gen Agent Trust Hub on Mar 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONNO_CODE
Full Analysis
  • [COMMAND_EXECUTION]: The documentation describes tools and structures that allow agents to execute arbitrary shell commands, such as the Bash tool and MCP server startCommand and installCommand configurations.- [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of external software through npmDependencies, pipDependencies, and npx commands for fetching MCP servers like @anthropic-ai/mcp-server-filesystem.- [PROMPT_INJECTION]: The architecture relies on interpolating instructions into prompts for commands and subagents, creating a surface for indirect prompt injection when processing untrusted data.
  • Ingestion points: Agent memory files (agentMDFile), workspace files, and outputs from tools or MCP servers.
  • Boundary markers: The documentation does not explicitly detail the use of delimiters or 'ignore' instructions to isolate data from instructions.
  • Capability inventory: Includes Bash tool, file modification (Write, Edit), and package installation.
  • Sanitization: No specific sanitization or validation logic is presented for handling external content within prompts.- [NO_CODE]: The skill consists entirely of markdown documentation and type definitions, containing no executable scripts or code to analyze directly.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 15, 2026, 04:12 AM