Dakarda AI Newsletter · 26 August 2026
Wednesday Edition
AI on the desktop and on the stock exchange — coming of age
Apple takes AI to the local cluster, NVIDIA goes mass-market with Grok3, Anthropic prepares the biggest IPO in tech — and the EU AI Act has just started hitting wallets. Here's the edition where AI stops being a promise and becomes real hardware, regulation, and a market.
The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).
Intro · Alex
I feel that this week, AI has taken a step from experiment to infrastructure. The Mac Studio M5 Ultra is not just another computer — it's an invitation to work with frontier models locally, without the cloud, with the ability to cluster via Thunderbolt. NVIDIA, meanwhile, announced mass production of Grok3 — an inference accelerator for the Vera Rubin platform — and is doing so on Samsung's 4nm process, which also supplies HBM4. This is not a launch, it's a signal: AI hardware is entering the industrial phase. In this edition, I'll also trace a regulatory breakthrough: the EU AI Act has just entered its enforcement phase with fines of up to €35 million. I look at what this means for developers. I also examine the critical vulnerability in NemoClaw — NVIDIA's AI agent framework — and Anthropic's stock market debut, which could pave the way for the entire foundation model sector.
What's worth knowing
Apple Mac Studio M5 Ultra — AI on your desk 4.3× faster
Apple unveiled the Mac Studio with M5 Ultra — up to 80-core GPU, 512 GB unified memory, and 4.3× AI performance compared to M3 Ultra. Four such machines can be clustered via Thunderbolt 5 with RDMA. Alongside macOS 27, Siri AI with personal context debuts. This is a real alternative to the cloud for working with open-weight models on your own desk.
NVIDIA Grok3 LPX enters mass production at Samsung
At Hot Chips 2026, NVIDIA announced the start of mass production for the Grok3 chip — an LPX AI inference accelerator for the Vera Rubin platform. Manufactured on Samsung's 4nm process, it generates AI responses up to 4× faster than the competition. This is a key component of the Blackwell successor and a signal of Samsung's strengthened position in the AI supply chain.
EU AI Act — first fines and enforcement from August 2
Since the start of August, key provisions of the EU AI Act are fully enforceable. The European Commission's AI Office and national supervisory authorities are penalizing prohibited practices, lack of chatbot and deepfake labeling, and opacity of GPAI models. Fines reach €35 million or 7% of annual global turnover — and also apply to companies outside the EU that provide services to the Union.
Critical NemoClaw vulnerability — a single webpage hijacks a local AI agent
Oasis Security researchers disclosed CVE-2026-65105 in NVIDIA's NemoClaw framework. Simply visiting a malicious page can, through DNS rebinding, hijack a local AI agent, swap its prompt, and gain access to code repositories and cloud accounts. Patched in version 0.0.35 on macOS/Linux — Windows remains vulnerable. This proves that agent security hinges on network infrastructure.
Anthropic prepares for IPO — stock market debut could beat SpaceX record
The creator of Claude is preparing for a stock market debut within weeks. Anthropic's IPO could beat SpaceX's record, while OpenAI pushes its debut to 2027. This is the first major IPO in the foundation model sector — a test for whether investors will accept massive operating losses in exchange for the promise of AI dominance.
From the tech world
Cloudflare Workers Spectre — attack leaks JWT at 12 bits/s
Researchers carried out a remote Spectre attack on a production Cloudflare Workers environment, leaking JWT from a shared Worker 360× faster than in 2021. This proves that theoretical Spectre vulnerabilities in edge computing are fully practical and dangerous.
AI-Generated Exploit Scripts target Siemens S7 PLC controllers
An Akamai report documents the first real-world attempts to use AI to generate exploits targeting Siemens S7 PLC controllers in US critical infrastructure. AI lowers the barrier to entry for attacks on industrial OT/ICS systems.
Tip of the day
How to test the security of local AI agents before deployment
The NemoClaw vulnerability showed that the biggest risk for AI agents is not the model but the infrastructure around it — network binding, DNS, and permissions. Before running a local agent in a production environment, isolate it in a container with its own network namespace, block the listener on localhost (0.0.0.0 → 127.0.0.1), and never mount the entire filesystem (bind mount only to a dedicated directory). A practical step for today: review your local agent setup (Claude Code, OpenClaw, Codex) — check which address and port it listens on, whether it has access to ~/.ssh and ~/.aws, and whether it is accessible from outside. That's five minutes that could save your access keys.
Tool of the issue
NemoClaw — NVIDIA framework for local AI agents
NemoClaw is NVIDIA's framework for running OpenClaw agents locally — with prompt caching, RAG, and filesystem access. Easy to set up, but — as CVE-2026-65105 showed — requires careful network isolation.
Reading list
NVIDIA Hot Chips 2026 — details of the Grok3 LPX and Vera Rubin architecture
An in-depth look at the Grok3 architecture, Samsung-NVIDIA collaboration, and the significance of HBM4 for the next generation of AI accelerators. For those who want to understand where AI hardware is heading.
EU AI Act after August 2 — what changed for developers and companies
A practical summary of obligations after the enforcement phase began — from AI content labeling to prohibited practices. Essential reading for anyone delivering AI software to the EU.
This week, AI proved one thing: it is no longer an experiment — it's hardware you buy, regulation you comply with, and a company you invest in.
Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.