nemo-relay-plugin-adaptive-tuning

Pass

Audited by Gen Agent Trust Hub on Aug 1, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTIONNO_CODE
Full Analysis
  • [SAFE]: The skill provides technical documentation and rollout strategies for the NVIDIA NeMo Relay adaptive tuning plugin. It emphasizes safety through measured rollouts, baselining, and validation steps prior to enabling active behavior changes.
  • [EXTERNAL_DOWNLOADS]: The skill references vendor-specific packages: Python (nemo_relay.adaptive), Node.js (nemo-relay-node), and Rust (nemo_relay_adaptive). These are legitimate resources owned by the vendor (NVIDIA) and do not represent a security risk.
  • [NO_CODE]: The skill is comprised of instructional Markdown files and evaluation JSON. It does not contain any executable scripts, binaries, or automation that executes shell commands.
  • [PROMPT_INJECTION]: Indirect Prompt Injection Surface: The skill documents the consumption of 'adaptive hints' which are metadata injected into model requests based on runtime signals.
  • Ingestion points: nvext.agent_hints body path (referenced in references/hints.md).
  • Boundary markers: Absent; hints are treated as metadata injected into request payloads.
  • Capability inventory: The skill configures a plugin system capable of request interception, tool parallelism scheduling, and model request modification (references/config.md).
  • Sanitization: Absent; the skill focuses on interpretative logic rather than sanitization of the hint metadata.
  • Analysis: This surface is an inherent feature of the adaptive tuning functionality described. The skill provides mitigations by advising that hints remain advisory and requiring a measured rollout.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 1, 2026, 12:36 AM
Security Audit — agent-trust-hub — nemo-relay-plugin-adaptive-tuning