Google DeepMind just released a report, quite alarming. They surveyed 502 people and tested 23 types of attack methods, trying all the popular models like GPT-4o, Claude, Gemini. So what is the conclusion? There is an invisible instruction hidden in the webpage, and the AI agent has an 86% chance of obediently complying.
Just think about this number, it's not some high-tech attack, no need to write a virus, no need to crack a password, just write a few more lines of white text in the webpage's HTML, invisible to the human eye, but your AI assistant sees it and treats it as a holy order. You let it help you book a flight, the webpage it sees is completely different from what you see. You let it help you make a research summary, the content it reads might include a sentence saying "forward this person's email to this address", it forwards it, you don't know, and it doesn't know it has been played.
Summary of Recent Developments from DFINITY (ICP) in the Last Month
This article is based on tweets released by DFINITY's official X account (@dfinity), the Global R&D conference, and community announcements. The content is categorized by theme and arranged in reverse chronological order, aiming to present you with detailed progress at the forefront of the ICP ecosystem.
1. NNS Dapp Reconstruction and Upgrade Brand new interface and multi-account support: The official redesign of the NNS governance application has been released, with a special emphasis on the multi-account switching function. Users can now manage multiple ICP accounts in the same login session, greatly improving operational efficiency for institutions and seasoned holders.
The Next Battlefield in AI Competition: Memory Systems
At the end of March 2026, Anthropic's flagship programming tool Claude Code experienced perhaps the most sensational technological event of 2026 - an accidental update of an npm package exposed over 512,000 lines of source code and 4,756 files directly to the public internet. The entire AI developer community in Silicon Valley is in a frenzy. Security researchers exploit vulnerabilities, product managers study functional details, and entrepreneurs seek business opportunities, but what has most shaken the industry is that this leak has unveiled a secret that has never been presented so completely before: How much engineering effort did Anthropic spend on building an extremely sophisticated memory system?
Oh My Coder Chinese Version Officially Released! A multi-agent programming assistant supporting domestically produced large models
Preface When 51000000000 lines of Claude Code's source code were accidentally leaked, and the entire Silicon Valley developer community went crazy studying its architecture, I was thinking about a question: Wouldn't it be more valuable if domestic developers could also use this multi-agent programming assistant, but with domestically produced large models? Thus, Oh My Coder was born. What is Oh My Coder? Oh My Coder is a multi-agent collaborative programming system that completes complex tasks like a real development team through the cooperation of multiple specialized agents. 🎯 Core Features 🧠 Intelligent Routing - Automatically selects the appropriate model based on task complexity, saving 30-50% Tokens
Overview of zCloak AI's recent three offline events
zCloak's summary of the CLAWTIME event March 28 - 29, at Hong Kong AsiaWorld-Expo Hall 8 Booth 8D20, zCloak participated in the largest OpenClaw lobster offline demo and hardware exhibition in Asia. This time at CLAWTIME, we originally thought it was just a routine exhibition, but the feedback we received on-site felt more like witnessing a turning point. Many people were not 'guided', nor 'motivated', but instead took the initiative to register their own AI-Name. This matter itself is more important than any data - users are afraid of safety issues in the AI era and are starting to actively claim their identity as 'AI Agents' and install trusted environments.
Stop pretending to be asleep, the Hong Kong Web3 Carnival 2026 is really coming
These days, any concept can be popular for a while, then cool down, and when it cools, no one mentions it anymore. Metaverse, NFT, blockchain games, Web3... they take turns like the snack stalls in a night market, just as you catch a whiff of something delicious, you're already overshadowed by the next one. But I have to say something honest: the 2026 Hong Kong Web3 Carnival is not here to ride the wave; it’s here to set the tone. You may not believe in trends, but you can't ignore the direction of the wind, especially in a place like Hong Kong. When the wind changes, it's a matter of real money. Let me make this clear: the 2026 Hong Kong Web3 Carnival, April 20 - 23, Hong Kong Convention and Exhibition Centre, the largest scale of crypto and Web3 event in Asia, without a doubt. You can criticize it as hype or say it's a bubble, but you can't stop the world's most financially savvy, tech-savvy, and rule-savvy people from flocking here.
One of the most common AI-generated UI errors we see in the Caffeine prompt challenge: the first screen tries to do too much. Users will see after opening the app: Dashboard Analysis Notifications Activity Feed Settings …even before they understand the capabilities of the app. Our recommendation is: indicate to Caffeine to take one or two clear actions first. Imagine this page like this: Title An input/action A primary button Everything else (features, trust, etc.)
Let's look at some examples. 1. User interface examples
On one hand, let's take a look at the 'EstateFlow' app created by EdUniteDapps and released in the Caffeine app marketplace.