agent-swarm-workflow

Pass

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The orchestration workflow requires agents to ingest and interpret instructions from external project files and communication channels.
  • Ingestion points: Agents are explicitly instructed to read and follow the content of AGENTS.md and README.md as part of their initial marching orders and subsequent implementation loops. They also monitor and respond to messages via the Agent Mail server.
  • Boundary markers: The prompt templates provided do not utilize delimiters (such as XML tags or triple quotes) to separate the agent's instructions from the data being processed, nor do they include warnings to ignore instructions found within that data.
  • Capability inventory: The agents are granted capabilities to modify source code, perform git commits, and execute local management tools.
  • Sanitization: No mechanisms for sanitizing input or validating the source of external data are defined within the skill's instructions.
  • [COMMAND_EXECUTION]: The skill documentation includes and encourages the use of local command-line tools for agent management and task prioritization.
  • Evidence: The skill references the use of ntm (Named Tmux Manager) for spawning agent sessions and bv (Bead Viewer) for automated task triage and planning. It also mentions a local shell script ~/projects/mcp_agent_mail/scripts/run_server_with_token.sh for managing the coordination server.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 4, 2026, 09:51 AM