Solutions  ·  2026-08-25

NVIDIA NemoClaw v0.0.114 hardens agentic MCP tool-call and sandbox lifecycle controls

SolutionsMedium impactGlobal
NVIDIA shipped NemoClaw v0.0.114 (Aug 23, 2026) adding deterministic read-only MCP tool calls with mutation rejection, credential-shaped-output redaction, destructive-operation warnings, sealed-configuration recovery with integrity verification, and fail-closed installer/image-provenance checks.
Converts several agentic-AI risk mitigations (excessive tool authority, secret leakage via tool output, unverified supply chain) into enforceable platform controls rather than system-prompt guidance, following the same product's v0.0.109 release a week earlier.
Teams building agent runtimes on NVIDIA NemoClaw/agentic sandboxes should adopt the read/write tool separation and image-provenance checks as a baseline control pattern.
NVIDIA NemoClaw Release Notes
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →