arbor

Warn

Audited by Gen Agent Trust Hub on Sep 3, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The documentation in references/arbor-upstream.md suggests that users clone and install a third-party tool from github.com/RUC-NLPIR/Arbor. This repository is not verified as a trusted source within the internal safety framework.\n- [INDIRECT_PROMPT_INJECTION]: The skill's workflow involves ingesting and processing data from external sources, such as artifact code and evaluator outputs, which serves as a surface for indirect prompt injection.\n
  • Ingestion points: The coordinator agent processes research insights and artifacts produced by executor sub-agents, as well as logs from development and test evaluation scripts.\n
  • Boundary markers: No explicit delimiters or safety instructions are defined to separate untrusted data from the agent's decision-making logic in the scripts/tree.py manager.\n
  • Capability inventory: The skill utilizes powerful tools including Bash, Agent (for sub-agent orchestration), Read, Write, and Edit.\n
  • Sanitization: The system lacks validation or sanitization of research insights before they are abstracted and propagated up the hypothesis tree.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 3, 2026, 04:56 AM
Security Audit — agent-trust-hub — arbor