fleet
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents an orchestration pattern where a parent agent monitors and consumes output from multiple child agents via
agent-deck session output. This creates a potential surface for indirect prompt injection if child agents process untrusted external data (such as third-party codebases) and return it to the parent. - Ingestion points: The parent agent collects data via the
agent-deck session output <child-id> --jsoncommand as documented in SKILL.md. - Boundary markers: The skill mentions a completion sentinel (
===AGENTDECK_DONE===) for status tracking, but does not provide specific delimiters or instructions to sanitize the natural language content returned by child agents. - Capability inventory: The parent agent has capabilities to launch new sessions (
agent-deck launch), send messages (agent-deck session send), and approve actions (agent-deck session approve). - Sanitization: No specific sanitization or validation of the content of child agent responses is described in the workflow.
Audit Metadata