sub-agent-orchestrator

Pass

Audited by Gen Agent Trust Hub on Apr 10, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The orchestrator design chains outputs from one agent into the prompts of another using variable interpolation (e.g., {{steps.research.output}}). This enables indirect prompt injection where untrusted data retrieved by one agent can manipulate the instructions given to subsequent agents. * Ingestion points: Prompts and step inputs defined in the workflow YAML. * Boundary markers: None present; data is directly concatenated. * Capability inventory: The skill utilizes the Agent, Read, Write, and Bash tools. * Sanitization: No validation or escaping is performed on interpolated variables.
  • [COMMAND_EXECUTION]: The workflow engine supports an 'eval' directive for conditional logic and grants sub-agents access to the 'Bash' tool. The evaluation of dynamic expressions containing data from previous steps poses a risk of command injection.
  • [COMMAND_EXECUTION]: Inconsistency detected between the skill's allowed-tools manifest (Read, Write, Agent, Bash) and the tools requested in the example workflow definitions (e.g., WebSearch), which may lead to unexpected behavior or policy bypass attempts.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 10, 2026, 05:26 PM