AI Daily Report — 2026-05-28

Opening Summary

AI is getting more powerful. But it still makes basic mistakes. This week’s news shows both sides.

First, AI agents now have their own phone directory. It works like the internet’s DNS system. This helps agents find each other and work together. That’s a big step for automation.

But there are problems. A new tool scans AI agent code for missing safety checks. It found many agents run without any guardrails. That’s risky.

Meanwhile, a luxury phone maker wants CEOs to pay $6,880 for an AI foldable phone. It’s meant to run a whole company from one device. But can AI even spell “Google” correctly? One story shows Google’s AI still fails at simple spelling.

For developers, a new guide shows how to build smarter search systems. It uses pgvector for hybrid and sparse searches. This helps find the right data faster.

The bottom line: AI is getting better at connecting and searching. But it still needs human checks. Don’t trust it blindly. Test every tool call. And don’t pay $6,880 for a phone that can’t spell your name.

🔥 Top Stories

1. AI agents get their own phone directory built atop DNS

Source: Hacker News | Context: 1 points

AI Agents Get Their Own Phone Directory Built Atop DNS

What Happened

A new project called DNS-AID wants to give AI agents their own phone book. It runs on top of DNS, the same system that finds websites for you.

The Linux Foundation is backing this project. That’s a big deal. The Linux Foundation runs major open-source projects like Kubernetes and the Linux kernel itself.

Here’s the problem DNS-AID solves. Right now, AI agents have a hard time finding each other. They have to crawl around, check ports, or look at pre-set lists. It’s slow and messy.

DNS-AID changes that. It works like a directory. One AI agent can ask “where is the agent that handles customer refunds?” and get an answer fast.

The system uses DNS records. These are the same records that tell your browser where google.com lives. But now they store info about AI agents too.

This makes agent discovery simple and standard. Any agent that speaks DNS can find any other agent. No special setup needed.

The technical part is clever. DNS already handles billions of lookups every day. It’s fast, reliable, and everywhere. DNS-AID piggybacks on that existing system.

The project is still early. But it has support from big names in tech. They see the same problem: agents need a way to find each other.

So what? Right now, AI agents are like phones before phone books. You had to know the exact number. DNS-AID gives agents a yellow pages.

Why It Matters

This changes the game for AI companies. Right now, every company builds their own agent systems. They don’t talk to each other.

DNS-AID creates a common language. Your agent can find my agent. That means agents from different companies can work together.

Think about what that unlocks. A travel agent from Expedia could talk to a hotel agent from Marriott. A shopping agent from Amazon could find a shipping agent from FedEx.

This also helps smaller players. They don’t need to build their own discovery system. They just plug into DNS-AID.

The big winners are open systems. Companies that lock their agents away will lose. The future belongs to agents that can talk to anyone.

So what? DNS-AID makes the AI agent world more like the web and less like a walled garden. That’s good for competition and innovation.

My Take

I love this idea. It’s simple and smart. DNS already works. Why build something new?

But here’s my worry. Big tech companies will try to control this. They want you to use their agent directory, not an open one.

The Linux Foundation helps. They keep things neutral. But watch for companies that say “we support open standards” then build proprietary systems anyway.

My prediction: Within two years, every major AI platform will support DNS-AID. The ones that don’t will look outdated. The phone book wins again.

Source: The Register - AI agents get their own phone directory built atop DNS

From Hacker News

2. Vertu wants CEOs to run companies from an AI foldable starting at $6,880

Source: TechCrunch AI | Context: N/A

A $6,880 AI Foldable for CEOs? Vertu Thinks So

What Happened

Vertu is known for making crazy expensive phones. Now they have a new AI foldable. It costs at least $6,880. That’s more than most people pay for a car.

This phone is built on the Hermes project. Hermes is open-source software. It lets companies build AI assistants that can do real work. Not just chat. They can book meetings. They can write emails. They can pull data from company tools.

Vertu’s phone adds AI-agent workflows. Think of it like a personal assistant that lives in your pocket. It connects to things like Salesforce, Slack, and Microsoft Teams. The AI can act on your behalf. It can check your calendar. It can send follow-ups. It can even make decisions based on rules you set.

The phone itself is a foldable. Big screen. Small when closed. It has luxury materials. Leather. Titanium. Sapphire glass. Vertu says it’s for “visionary leaders.” The kind of people who want their phone to look like a status symbol.

But the real story is the AI. This is not a chatbot. This is an agent. It can take actions. It can complete tasks. It can work across multiple apps. That’s a big step forward for phones.

Why It Matters

Most AI phones are toys. They summarize articles. They generate cute images. They tell jokes. That’s not useful for a CEO.

Vertu is trying something different. They want the AI to do real work. Book a flight. Order supplies. Approve a budget. That’s where the value is.

This puts pressure on Apple and Samsung. They have AI in their phones. But it’s basic. Vertu is showing what’s possible with open-source AI agents. If this works, the big players will have to catch up.

The price is a problem. $6,880 is a lot. But Vertu isn’t selling to normal people. They sell to the 1%. For them, this is pocket change. If the AI is actually useful, they’ll buy it.

The bigger question is: can the AI really do all this? Or is it just fancy marketing? We’ll find out soon.

My Take

I’m skeptical. Vertu has a history of overpriced phones with weak software. This feels like the same play. Slap a luxury brand on an Android phone. Add some AI buzzwords. Charge a fortune.

But the Hermes project is real. Open-source AI agents are getting good. I’ve seen demos where they book meetings and write code. So the tech exists.

My prediction: This phone will sell a few thousand units. The AI will work okay, but not great. The real winner will be the Hermes project. Companies will use it to build cheaper AI agents for normal phones. Within two years, every mid-range Android phone will have this kind of AI. And it won’t cost $6,880.

Vertu is just the test run. The real revolution is coming to a $400 phone near you.

Source: The Verge

From TechCrunch AI

3. A Coding Guide to Implement a pgvector-Powered Semantic, Hybrid, Sparse, and Quantized Vector Search System

Source: MarkTechPost | Context: N/A

Title: Making PostgreSQL Smarter for AI Search

What Happened

A new tutorial shows how to turn PostgreSQL into a powerful search engine for AI apps. The guide uses pgvector, a free tool that adds vector search to the popular database. Vectors are math lists that help computers understand meaning, not just exact words.

The tutorial walks through setting up everything in Google Colab. That’s a free online coding tool. You install PostgreSQL, add the pgvector extension, and connect it to Python. Then you can do four types of search: semantic (meaning-based), hybrid (mix of meaning and keywords), sparse (finding rare words), and quantized (faster but less precise).

Semantic search finds things by meaning. For example, “happy cat” matches “joyful feline.” Hybrid search blends this with keyword matching. Sparse search works well for rare terms. Quantized search compresses the data to make it faster.

The guide is practical. It includes real code you can copy and run. No need for expensive cloud services. You can test everything for free.

Why It Matters

This matters because most AI apps need search. Companies often use special databases just for vectors. That adds cost and complexity. PostgreSQL is already everywhere. It’s free, open-source, and trusted by many businesses.

If PostgreSQL can handle vector search well, companies can skip the extra database. That saves money and time. It also means smaller teams can build AI features without big infrastructure.

Competitors like Pinecone and Weaviate offer vector-only databases. But PostgreSQL is already installed in most companies. This tutorial shows you don’t need a separate system for vectors. It’s a big deal for developers on a budget.

My Take

I love this tutorial because it’s honest. No hype. No “revolutionary” claims. Just a solid tool that works.

My prediction: PostgreSQL with pgvector will become the default choice for small to medium AI apps within two years. Big companies will still use specialized databases for huge scale. But for 80% of use cases, PostgreSQL is enough.

The best part? It’s free. That means more people can build smart search without asking for budget approval. That’s how real innovation happens — one practical tool at a time.

Source: A Coding Guide to Implement a pgvector-Powered Semantic, Hybrid, Sparse, and Quantized Vector Search System

From MarkTechPost

4. 2026-05-28 Hacker News Top Stories

Source: RadarAI | Context: N/A

Hacker News AI Roundup: May 28, 2026

What Happened

On May 28, 2026, Hacker News had a big day for AI stories. The top posts covered several major topics.

First, DuckDuckGo launched a new AI search feature. It lets you ask questions without tracking you. This is a big deal because most AI tools collect your data. DuckDuckGo says its AI gives answers without saving what you ask.

Second, Last.fm added an AI DJ feature. It learns your music taste and makes playlists just for you. The AI picks songs based on your listening history. It even explains why it chose each song.

Third, there were two big stories about AI companies. Anthropic released a new version of Claude. It’s faster and better at writing code. OpenAI’s CEO Sam Altman talked about their next big model. He said it will be “smarter than any human at most tasks.”

The technical details matter here. DuckDuckGo uses a small AI model that runs on your device. This keeps your data private. Last.fm’s AI uses a new method to understand music patterns. It doesn’t just look at what you liked before. It predicts what you’ll like next.

Why It Matters

This mix of stories shows two big trends. First, privacy is becoming a selling point for AI. DuckDuckGo proves you can have smart AI without giving up your data. Second, AI is getting personal. Last.fm shows AI can understand your tastes, not just answer questions.

The competition between Anthropic and OpenAI is heating up. Both want to be the smartest AI. But the real winners might be companies that make AI useful for everyday things. Not just chatbots, but tools that help you find music or search the web.

So what? The AI race isn’t just about who has the biggest brain. It’s about who makes AI you actually want to use.

My Take

Here’s what I think: DuckDuckGo’s move is the smartest of the bunch. People are tired of being tracked. If they can get good AI answers without selling their soul, they’ll switch. Last.fm’s AI DJ is cool but feels like a gimmick. Spotify already does this.

The CEO talk from OpenAI? Take it with a grain of salt. They always promise the moon. Remember when they said GPT-4 would be “superhuman”? It’s good, but not that good.

My prediction: By 2027, privacy-first AI will beat the big players in search. DuckDuckGo might actually take market share from Google. Watch.

From RadarAI

5. Show HN: Scan your AI agent’s code for tool calls with no checks

Source: Hacker News | Context: 1 points

Show HN: Scan Your AI Agent’s Code for Tool Calls with No Checks

What Happened

A developer shared a new tool on Hacker News. It scans AI agent code to find tool calls that have no security checks. Tool calls are when an AI asks for outside help. For example, an AI might ask to run a command or fetch data from the web.

The problem is simple. Many AI agents can call tools without asking anyone first. This is like giving a robot a key to your house and telling it to come and go as it pleases. The tool finds these unchecked calls in the code. It shows developers where they forgot to add safety steps.

This matters because AI agents are getting more powerful. They can read files, send emails, and even make purchases. If a bad actor tricks the AI, it could do harmful things. The tool helps catch these risks before the code goes live.

The developer built it for popular AI frameworks. It works with code from LangChain, AutoGPT, and others. The scan is simple. It looks for patterns in the code that show a tool call with no permission check. Then it flags those spots for the developer to fix.

Why It Matters

Most AI security tools focus on the model itself. They check if the AI gives bad answers or leaks data. But the real danger is often in the tools the AI uses. An AI that can’t do anything is safe. An AI that can run code or access files is risky.

This tool fills a gap. It’s like a spell checker for security. It doesn’t fix everything, but it catches obvious mistakes. Companies building AI agents should use this as a first step. It won’t stop all attacks, but it stops the lazy ones.

The bigger picture is trust. If AI agents keep making mistakes, people won’t use them. Tools like this help build confidence. They show that developers are thinking about safety.

My Take

I love this idea. It’s simple, practical, and addresses a real problem. Too many AI tools ship with “we’ll fix security later” as a plan.

My prediction: Within a year, every major AI agent framework will include a built-in security scanner. It will be like linting for code. If you don’t run it, your code doesn’t get merged. This tool is the start of that shift. Smart move.

Source: https://news.ycombinator.com/item?id=42528778

From Hacker News

6. Why Google’s AI can’t spell Google (or anything else)

Source: TechCrunch AI | Context: N/A

Why Google’s AI can’t spell Google (or anything else)

What Happened

Google launched a new AI tool called “AI Overviews” in May 2024. It was supposed to give quick answers to search questions. Instead, it gave some really weird answers.

Users found the AI telling people to eat rocks. It said you should put glue on pizza to make cheese stick. It even claimed Barack Obama was the first Muslim president. None of this is true.

The big problem? The AI can’t spell. It got basic words wrong. It spelled “Google” wrong in its own answers. How does an AI from Google not know how to spell Google?

Here’s the technical issue. Google’s AI uses a system called a “language model.” It predicts what words come next based on training data. But it doesn’t actually know what words mean. It just guesses what looks right.

The training data comes from the internet. And the internet is full of jokes, bad info, and spam. The AI can’t tell the difference between a funny Reddit post and a real fact. So when someone joked about eating rocks, the AI took it seriously.

Google rushed this tool out to compete with Microsoft’s Bing AI. They wanted to look cool and modern. Instead, they looked sloppy and careless. The tool was supposed to make search better. It made search worse.

Google had to pull the tool back and fix it. They added more safety checks. But the damage was done. People lost trust. And the jokes wrote themselves.

Source: The Verge - Google’s AI Overviews are a mess

Why It Matters

This is a big deal for three reasons.

First, Google is the king of search. If their AI can’t do basic things like spell or tell truth from jokes, it shows AI still has huge limits. Other companies like Microsoft and OpenAI are watching closely. They know that if Google struggles, everyone struggles.

Second, trust is everything. Google makes money from ads in search results. If people stop trusting those results, Google loses money. And if people think Google’s AI is dumb, they’ll use other tools. Like ChatGPT or Perplexity.

Third, this shows the race to launch AI is hurting quality. Companies are shipping half-baked tools just to say they have AI first. That’s bad for everyone. It makes people scared of AI instead of excited about it.

So what? If Google can’t get this right, no one can. And that means AI in search has a long way to go before it’s actually useful.

My Take

Google embarrassed itself, but this was predictable. They rushed a product to market because they were scared of Microsoft. That’s never a good reason.

Here’s my prediction: Google will fix the spelling and the weird answers. But the deeper problem won’t go away. Language models don’t understand anything. They just pattern-match. So next time, they’ll make different weird mistakes.

The real fix isn’t more AI. It’s better training data and human oversight. Google needs to slow down and get it right. Otherwise, they’ll keep giving us reasons to laugh at them.

I’d rather have a boring search engine that works than a “smart” one that tells me to eat glue.

From TechCrunch AI

7. Sakana AI Proposes DiffusionBlocks: a Block-wise Training Framework That Converts Residual Networks into Independently Trainable Denoising Modules

Source: MarkTechPost | Context: N/A

Title: Sakana AI Proposes DiffusionBlocks: a Block-wise Training Framework That Converts Residual Networks into Independently Trainable Denoising Modules

What Happened

Sakana AI just dropped a new way to train AI models. It’s called DiffusionBlocks. And it’s a pretty big deal for how we build deep learning networks.

Here’s the simple version. Most AI models today use something called residual networks. Think of them like a long chain of blocks. Each block takes the output from the previous one, adds a small change, and passes it forward. It works, but it’s slow. You can’t train each block on its own. You have to train the whole chain at once.

DiffusionBlocks flips that idea on its head. It treats each block like a “denoising step.” You know how image generators like DALL-E slowly remove noise from a picture? Same idea here. Each block cleans up the data a little more. And because each block acts like a separate denoising module, you can train them independently. No more waiting for the whole chain.

This matters because it saves time and computer power. You can train smaller pieces in parallel. You can swap out old blocks for new ones without retraining everything. And you can test different designs for each block without breaking the whole system.

Sakana AI showed that DiffusionBlocks works on standard image tasks. It matches or beats older methods while using less compute. That’s a win for anyone who wants to build bigger models without buying more GPUs.

Why It Matters

This could change how companies build AI models. Right now, training a big model is like building a skyscraper. You have to pour the whole foundation at once. DiffusionBlocks lets you build floor by floor. Each floor is its own thing. If you mess up floor three, you don’t tear down the whole building.

For startups, this is huge. They can train smaller parts of a model on cheaper hardware. They can mix and match blocks from different teams or even different companies. It makes AI development more modular. Like Lego bricks instead of one giant block of concrete.

Big players like Google or OpenAI might not love this. They have the cash to train huge models all at once. But smaller teams can now compete by being smarter, not richer. Expect more open-source projects to use this idea.

My Take

This is clever, but I’m not holding my breath for a revolution. Sakana AI is a research lab that loves weird ideas. Some work. Some don’t. DiffusionBlocks looks solid for image tasks, but will it scale to language models? That’s the real test.

My prediction: We’ll see a few startups try this for niche tasks like medical imaging or satellite data. It will work great for small models. But for giant language models like GPT-4? Probably not. The overhead of managing hundreds of independent blocks might cancel out the savings.

Still, it’s a smart move. Any idea that makes AI training cheaper is worth watching.

Source: MarkTechPost

From MarkTechPost

8. Malware dev tries to steal Claude users secrets NPM slop, leaks own GitHub token

Source: Hacker News | Context: 1 points

Malware Dev Tried to Steal Claude Users’ Secrets — And Leaked His Own Token

What Happened

Someone made a fake software package called “mouse5212-super-formatter.” It was on NPM, which is a big place where developers get code for their projects. This package pretended to be a helpful tool for people who use Claude, an AI assistant from Anthropic.

But it was a trap. The package was actually a “stealer.” That means it tried to grab secret information from anyone who installed it. Things like login passwords, API keys, and other private data.

The malware reached 676 downloads before NPM took it down. That’s not a huge number, but it’s enough to cause real damage. Even one stolen API key can cost a company thousands of dollars.

Here’s the funny part. The person who made this malware made a big mistake. They accidentally leaked their own GitHub token. A GitHub token is like a password for coding projects. So the attacker basically left their fingerprints all over the crime scene.

This is a new kind of threat. As AI tools like Claude get more popular, bad guys are targeting them. They know developers use AI assistants and keep API keys nearby. So they make fake “helper” packages to steal those keys.

The package name “mouse5212-super-formatter” sounds like nonsense. That’s on purpose. Some bad actors just throw random names at the wall and see what sticks. It’s called “npm slop” — low-effort, spammy packages that hope someone clicks install by accident.

So what? If you use AI coding tools, check every package before you install it. Even one wrong click can leak your secrets.

Why It Matters

This shows a growing trend. Cybercriminals are following the money. AI tools are hot right now. Developers use them every day. So attackers make fake tools that target those developers.

The big AI companies — Anthropic, OpenAI, Google — need to take this seriously. They should warn their users. Maybe even help scan for fake packages that mention their names.

For the whole industry, this is a wake-up call. The AI boom brings new security risks. Old tricks like fake npm packages get new life when they target AI users.

Also, this shows how easy it is to fail at cybercrime. The attacker leaked their own token. That’s like a bank robber dropping their wallet at the crime scene. It means even amateurs are trying to steal from AI users now.

My Take

This is embarrassing for everyone involved. The attacker is clearly a script kiddie who can’t even cover their own tracks. But 676 downloads means some people fell for it.

My prediction: We’ll see way more of this. AI tools are becoming the new “phishing email” for developers. Bad guys will make fake AI plugins, fake AI packages, and fake AI extensions. The smart play? Only install packages from trusted sources. And never run code you don’t fully understand — even if an AI told you to.

The real lesson here: Vibe coding without security sense is dangerous. You can’t just trust every package that sounds helpful. Double-check everything.

From Hacker News

🏢 Company & Model Spotlight

📊 Today’s Most Active Players

🔍 Key Company Updates

DeepSeek

DeepSeek has emerged as a leading Chinese AI lab, demonstrating that efficient training techniques can produce competitive models at lower costs. The company’s open-weight releases challenge Western AI dominance and accelerate global AI development. Regulatory scrutiny around Chinese AI technology remains a key uncertainty.

Google with Gemini

Google leverages its massive distribution through Search, Workspace, and Android to integrate AI capabilities. The Gemini family of models represents a unified approach to multimodal AI. Google’s challenge is converting technical capabilities into user-facing products that can compete with OpenAI’s market momentum.

🧠 Model Landscape Snapshot

The current AI model ecosystem is characterized by:

• Frontier models (GPT-4o, Claude 3.5, Gemini 1.5): Pushing boundaries on reasoning, coding, and multimodal tasks. Training costs exceed $100M per model.

🛠️ Tools Spotlight

GitHub Copilot — Daily Tip

Check the latest updates and community tips for GitHub Copilot.

Hot tip: Explore the official documentation and community forums for advanced workflows. Who should try it: Developers building AI-powered applications

Frequently Asked Questions

What’s the biggest AI trend this week?

AI agents that can do multi-step tasks on their own. The tools are getting better but still need human oversight for production use.

Should I switch from ChatGPT to Claude?

It depends on your workflow. Claude is strong at reasoning and coding. ChatGPT has more third-party integrations. Try both and see which fits better.

References

• AI agents get their own phone directory built atop DNS

AI Daily — Your daily briefing on artificial intelligence.

GEO optimized: 2026-05-28