Dakarda AI Newsletter · 15 July 2026
Wednesday Edition
The Great Firewall for AI and Other Boundaries
Wednesday brings a real breakthrough: China introduces the world's strictest law against AI impersonating humans, researchers find a way to poison agent memory with a single email, and someone runs a 744B model on a laptop without a GPU. Plus the EU AI Act hits 18 days before the deadline — things are heating up.
Intro · Alex
Wednesday, July 15, and China's regulations have just come into force, banning virtual family members and romantic AI partners for minors. This is the world's first comprehensive law of its kind — and a signal that emotional AI is no longer the Wild West. At the same time, researchers have unveiled MemGhost — an attack where a single email can implant false memories into an AI agent's memory. Sounds like science fiction? It's already a production-grade compromise vector. In today's edition, I've picked five stories that paint a clear picture: AI is entering the age of accountability (and regulation), but it's also opening new security fronts. Colibri proves frontier models can run locally — on a laptop, without a GPU. Google states plainly: 83% of companies lack the infrastructure for agents. And from August 2, all AI content in the EU must be labeled. We review it all with perspective and a practical eye.
What's worth knowing
China Bans AI Companions for Children — The World's Strictest Law
The "Interim Measures for Managing Anthropomorphic Interactive AI Services" have come into force — banning the offering of virtual family members or romantic partners to minors. Platforms like Toki (147 million users) must immediately adapt their services. This is the first law that explicitly says: AI cannot impersonate a loved one for children, setting a new global standard — hot on the heels of similar regulations in California and New York.
MemGhost — How to Poison an AI Agent's Memory with a Single Email
The OpenClaw team and independent researchers have published details of the MemGhost attack — a one-shot framework that installs false memories into an agent's persistent memory through a single crafted email. End-to-end effectiveness in lab tests: 87.5% across 56 scenarios. Fake data remains hidden and influences the agent's future decisions — without the user's knowledge. Target: HR, finance, and support assistants reading their inbox.
Colibri — 744 Billion Parameters on a Laptop Without a GPU
Developer JustVugg has released Colibri — a lightweight inference engine in pure C that runs GLM-5.2 (a 744B parameter model, MoE architecture) on consumer hardware with ~25 GB RAM and no GPU. The secret: streaming experts from an NVMe drive. Speed? 0.05–0.1 tokens per second — it's not a Claude replacement, but a proof of concept that shatters the dogma that massive models require data centers.
Google Cloud: 83% of Companies Need New Infrastructure for AI Agents
Google Cloud has published a report showing that 83% of organizations need to modernize their infrastructure before they can support agentic AI. The report accompanies Google's announcement of new investments in HPC and dedicated cloud solutions for autonomous agents. A signal for CTOs and CIOs: modernizing the stack (network, storage, GPU/latency) is not an option — it's a prerequisite for deploying agents in 2026–2027.
18 Days to Mandatory AI Content Labeling in the EU — Article 50 Takes Effect
From August 2, 2026, all AI-generated content (images, audio, video) must be marked in a machine-readable format — via invisible watermarks or C2PA content credentials. The European Commission is finalizing the Code of Practice. Risk: fines up to €35 million or 7% of global annual turnover. Applies to every company publishing AI content for audiences in the EU. This isn't the future — it's 18 days away.
From the tech world
Microsoft's Secure Boot Has Been Breakable for a Decade — Nobody Noticed
Researchers from ESET have discovered 11 old Microsoft-signed UEFI images (shims) — the oldest dating back to 2013 — that allow trivial bypass of Secure Boot. Microsoft failed to revoke the certificates for over a decade. This means every older device with Secure Boot (Windows and Linux) may have been vulnerable to firmware attacks — in a way accessible even to beginners.
Record Microsoft Patch: 622 Fixes, Including Two Zero-Days Under Active Attack
Microsoft has issued a record Patch Tuesday — 622 fixes, including CVE-2026-56164 (SharePoint Server) and CVE-2026-56155 (Active Directory Federation Services) already being exploited in campaigns detected by Mandiant and Google FLARE. More than triple the previous record. SharePoint Server 2016 and 2019 reached end of support on the same day.
Tip of the day
How to Secure Your AI Agent Against Memory Attacks
The biggest lesson from the MemGhost attack: an agent's persistent memory is not a vault — it's an attack vector. If your AI agent reads emails and makes decisions based on them (payment authorization, ticket responses), treat its memory file like an inbox — not a safe. Practical step: Implement a "confirm before executing" policy for all decisions with financial or legal consequences. The agent may remember fake invoices or instructions — but if every authorization requires a checkpoint with business logic (e.g., "does this account exist in the system?"), a single poisoned email won't be enough to cause harm. Additionally: regularly clean and validate the agent's persistent memory by cross-referencing it with external sources of truth.
Tool of the issue
Colibri — Run a 744B Model on Your Laptop (Really)
An open-source inference engine written in pure C, with no dependencies. Runs GLM-5.2 (744 billion parameters) on hardware with ~25 GB RAM and no GPU — thanks to streaming experts from an NVMe drive. Ideal for developers wanting to test frontier models locally and privacy enthusiasts.
Reading list
EU AI Act: A Practical Guide to AI Content Labeling
If you think Article 50 doesn't apply to you — read this. Concrete techniques (invisible watermarking vs C2PA), deadlines, and penalty risks. 18 days to the deadline, and implementation is not trivial.
Colibri from the Inside — Technical Review on Wavect
A detailed technical deep dive: how Colibri streams experts, why it works in 25 GB RAM, and what this means for the future of local inference. For those who want to understand, not just hear about it.
The most interesting thing this week isn't what AI can do — but where the world is drawing the lines: legal, security, infrastructural. And it's these boundaries that define what will be possible next.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.