agent-fungibility

Pass

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill's core workflow relies on agents reading and processing untrusted data from the local environment and communication channels, which could contain malicious instructions designed to hijack the agent's task execution.
  • Ingestion points: Agents are explicitly instructed to read and 'understand ALL' of AGENTS.md and README.md, and to 'promptly respond' to messages received via the MCP Agent Mail system.
  • Boundary markers: The provided initial prompt for the agents does not include delimiters (e.g., XML tags or triple backticks) or specific instructions to ignore embedded commands within the ingested documentation or messages.
  • Capability inventory: Agents are granted significant capabilities including code investigation mode, the ability to register with and use a messaging system, and the use of task-management tools (bv and ntm) to modify project state.
  • Sanitization: The workflow lacks defined sanitization or validation logic to filter out potentially harmful natural language instructions embedded in the processed project files or agent mail.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 4, 2026, 09:50 AM