agent-orchestrator
Pass
Audited by Gen Agent Trust Hub on Nov 12, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents architectural patterns for managing untrusted data ingestion (e.g., chat history, tool results) and specifically addresses risks like 'summary-injection' where models might inject instructions into their own compaction summaries. It recommends defensive measures such as using typed blocks for state, deterministic limits, and human-in-the-loop checkpoints.
- Ingestion Points: User questions, chat history, and external tool outputs processed by sub-agents.
- Boundary Markers: Recommends using delimiters and typed carryover blocks (e.g.,
task_focus,verified_work) to prevent context drift. - Capability Inventory: The architecture supports SQL execution, tool-calling loops, and memory retrieval.
- Sanitization: Suggests redaction, secret detection, and structured output validation (e.g., SQL schema checks).
- [DATA_EXFILTRATION]: The documentation identifies data exfiltration risks, such as agents splitting tokens to avoid secret scanning or using DNS egress to reach external chatbots. It provides remediation guidance, such as keeping credentials out of sandboxes and implementing resolver-level DNS filtering.
- [COMMAND_EXECUTION]: The skill describes tool-calling patterns but emphasizes the importance of read-only enforcement, DML blocking for SQL, and explicit authorization for external effects, rather than providing direct shell execution capabilities.
- [DYNAMIC_EXECUTION]: A code snippet in
references/patterns.mddemonstrates the use oftiktokenfor token estimation with a character-based fallback, representing standard library usage for context window management.
Audit Metadata