orchestrator
Pass
Audited by Gen Agent Trust Hub on Jul 15, 2026
Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill is configured to transmit repository data, code snippets, and prompts to external models including Grok (xAI) and Fable (Anthropic).
- Evidence: "A Grok dispatch can send its prompt, specification, code excerpts, and other repository content to xAI." (SKILL.md)
- Evidence: "A Fable consult can send its consult package and repository files inspected by read tools to Anthropic." (SKILL.md)
- Safeguards: The instructions require explicit user approval before the first transmission and strictly prohibit sending credentials, secrets, or unrelated proprietary content.
- [PROMPT_INJECTION]: The skill's architecture involves ingesting and acting upon instructions or code generated by external agents, presenting a risk of indirect prompt injection.
- Ingestion points: Specialist reports, implementation worker results (Grok), and advisor consults (Fable).
- Boundary markers: Uses read-only boundaries for advisors and partitioned file ownership for implementation workers.
- Capability inventory: The main agent possesses the authority to perform final synthesis, execute tests, commit changes, and push code to remote repositories.
- Sanitization: Requires "adversarial review" of every worker diff and main-seat verification before finalization.
Audit Metadata