We publish before the industry notices.

We found three security bugs in NVIDIA's NemoClaw. Then we fixed them -- on an Apple Mac Studio.

NemoClaw on Apple Silicon: Validating and Fixing NVIDIA's Enterprise Agent Security Stack on Consumer Hardware

Published in the Global Journal of Mobile Communication and Computing Technologies (MK Science Set) · DOI 10.63620/MKJAEAST.2026.1015

Adnan Tanveer (Addy) · Speedrun AI Labs, Sydney · April 13, 2026 · v2.1 · Also on SSRN 6587938

NVIDIA NemoClaw provides a six-layer enterprise security stack for autonomous AI agents running on the OpenClaw platform. The architecture was designed and announced alongside NVIDIA DGX hardware at GTC 2026. DGX Station systems list at US$80,000 to US$125,000. This paper validates the complete NVIDIA NemoClaw/OpenShell security stack on an Apple Mac Studio (M4 Max, 64 GB unified memory, A$4,849) via Docker Desktop. We identify three enforcement gaps in NVIDIA OpenShell v0.0.26 -- root-user bypass of Landlock via CAP_DAC_OVERRIDE, missing seccomp filters for AF_PACKET and AF_NETLINK, and a missing TLS CA certificate mount -- develop and publish fixes for all three gaps, achieving 6 of 6 security layers passing, and benchmark sandbox startup latency, memory overhead, and infrastructure cost, demonstrating that enterprise-grade agent security is achievable on Apple Silicon at approximately 4% of NVIDIA DGX Station pricing.

The first inside look at NVIDIA's NemoClaw. Published the day after GTC 2026.

Inside NemoClaw: An Architectural Analysis of NVIDIA's Enterprise Security Stack for Autonomous AI Agents

Published in the Global Journal of Mobile Communication and Computing Technologies (MK Science Set) · DOI 10.63620/MKJAEAST.2026.1014

Adnan Tanveer (Addy) · Speedrun AI Labs, Sydney · March 17, 2026, one day after GTC 2026 · v1.0 · Also on SSRN 6436138

NVIDIA NemoClaw, announced at GTC 2026 by Jensen Huang, is the first major vendor implementation of enterprise-grade security for OpenClaw autonomous AI agents. This paper presents the first detailed public decomposition of the NemoClaw and OpenShell security architecture across six layers: Landlock filesystem isolation, seccomp BPF syscall filtering, network namespace isolation, OPA/Rego policy enforcement, TLS inspection, and provider-agnostic inference routing, derived from direct inspection of both open-source codebases.

One AI model can't check its own work. So we convened a council of seven.

The Intelligence Council: Provider-Agnostic, Domain-Specialized Multi-Model Review for Autonomous AI Systems

Adnan Tanveer (Addy) · Speedrun AI Labs, Sydney · July 12, 2026 · v1.0 · Pending SSRN review

Most enterprise AI systems bind themselves to a single model, inheriting its blind spots, outages, and biases with no recourse -- a widely cited reason many AI pilots fail. This paper presents the Intelligence Council: a provider-agnostic pattern that convenes a panel of independent frontier models to review a plan, a code change, or a decision before it is acted upon, then synthesises their verdicts into consensus, dissent, and blockers. Its value is independence -- different vendors, different training, genuinely separate reasoning that a single model cannot supply when checking its own work. We document a live seven-model reference council, a four-domain taxonomy of specialised panels (Coding, SEO/Search, Cybersecurity, Intelligence) with benchmark-grounded rosters, a reproducible configuration failure mode, and a synthesis protocol. The pattern is model-agnostic and harness-agnostic by construction, embedding in any coding harness or agent framework.

One Apple Mac Mini. One NVIDIA Nemotron model. A fully autonomous AI agent -- replicated from PDFs.

Deploying Self-Improving AI Agents on Apple Silicon with OpenClaw and NVIDIA Nemotron

Adnan Tanveer (Addy) · Speedrun AI Labs, Sydney · April 13, 2026 · v2.1 · SSRN 6690499

Deploying autonomous AI agents in production requires extensive configuration of infrastructure, security policies, memory systems, scheduled tasks, and operational tooling. This paper documents the replication of a fully operational AI agent from one Apple Silicon machine to another using PDF specification documents executed by an AI deployment assistant. We present a nine-phase replication methodology that deploys a complete agent stack -- including nine custom skills, four scheduled cron jobs, a hardened self-improvement hook system, and local inference via NVIDIA Nemotron 3 Nano 30B on Ollama -- from specification documents alone; document a self-improvement architecture where agents log errors and corrections, promote recurring patterns to permanent identity files, and inject live learning summaries at every session bootstrap; and validate the deployment with a three-pass, 69-check verification audit achieving zero errors. The target agent operates on an Apple Mac Studio (M4 Max) in a regulated investment fintech environment, running marketing workloads with NVIDIA NemoClaw providing all six security layers.

The first production fix for OpenClaw's broken memory.

Building Fault-Tolerant Memory for OpenClaw AI Agents: A Three-Layer Architecture for the SQLITE_CANTOPEN Regression in 2026.3.x

Adnan Tanveer (Addy) · Speedrun AI Labs, Sydney · March 16, 2026 · v1.1 · SSRN 6436099

OpenClaw's default memory system exhibits recurring reliability failures: SQLite crashes, dirty indexes, and complete persistence loss during active sessions. This paper presents a three-layer fault-tolerant architecture combining pre-compaction memory flushing, QMD-based local hybrid search, and Mem0 self-hosted on Qdrant, validated in a production deployment on a Mac Mini that eliminated all documented failure modes while maintaining zero cloud dependency for search.