AI Summary

Market Overview

• AGIX posted a strong DTD gain of 3.13%, significantly outperforming the QQQ (+0.60%), S&P 500 (+0.26%), and Dow Jones (+0.09%) benchmarks, driven by broad strength in AI and semiconductor holdings.

• Among peer AI ETFs, IGV (+5.94%) and LRNZ (+6.14%) led, while SMH (+1.48%) and BOTZ (+1.12%) showed moderate gains, reflecting a mixed but generally positive sentiment across AI-related sectors.

• Top performing AGIX holdings included Twilio (+19.36%), Nebius Group (+14.46%), CoreWeave (+13.96%), Datadog (+12.19%), and Cloudflare (+11.99%), all benefiting from the agentic AI and enterprise software themes highlighted in news.

• Key stocks to watch: NVIDIA (+6.26%) continues to lead hardware; Arm (+15.73%) and Samsung (+10.09%) surged on chip demand; Oracle (+9.91%) and Salesforce (+9.68%) gained on enterprise AI adoption news.

News Highlights

1. Nvidia's GTC Taipei announcements (Agent Toolkit, Vera Rubin production, RTX Spark, DGX Station) reinforce its dominance in AI infrastructure and agentic AI, positively impacting AGIX's top holdings like NVDA, ARM, and Broadcom.

2. Intel’s edge AI wins and OpenVINO launch highlight competition in the edge/robot AI space, benefiting AGIX holdings like NVIDIA and indirectly supporting the broader AI hardware theme.

3. Anthropic's confidential IPO filing and $65B funding round signal strong demand for frontier AI models, likely boosting sentiment for AGIX infrastructure names like Microsoft, Alphabet, and Oracle.

4. Alphabet's $80B equity raise for AI infrastructure underscores massive capital expenditure plans, supporting AGIX's Alphabet and Microsoft holdings, while also validating the AI buildout theme.

5. Florida's lawsuit against OpenAI could introduce regulatory headwinds for the AI sector, but near-term impact on AGIX holdings appears limited given the focus on safety and data practices.

6. Dell's 88% revenue growth from AI servers and stock surge (+33% in a week) reflect strong hardware demand, benefiting AGIX holdings like NVIDIA and Broadcom.

7. SpaceX's IPO filing with water resource risk factors highlights infrastructure constraints, a theme that may affect AGIX's data center and energy-related holdings like Vertiv and Eaton.

8. Fingerprint's AI Assistant Detection and Merge's Agent Handler address security and governance challenges for AI agents, supporting AGIX's cybersecurity holdings like Palo Alto Networks and Zscaler.

RSS Highlights

1. Hackers exploited Meta's AI support bot to hijack high-profile Instagram accounts, highlighting security risks in AI-driven account recovery systems and the need for robust authentication.

2. A deep dive into the approximation formula exp(-x²) ≈ (1 + cos(sin(x) + x))/2 reveals its surprising accuracy over a wide interval, far outperforming Taylor series.

3. The IndieWeb movement gains traction with microformats, promoting decentralized, interoperable web standards as an alternative to centralized platforms.

4. Amazon launched AI-generated podcast-style audio clips for products, using two AI 'hosts' to discuss reviews and features, potentially changing how consumers research products.

5. The 1000th data breach recorded on Have I Been Pwned highlights that disclosure delays have worsened despite privacy regulations, with companies prioritizing legal liability over customer protection.

6. A new cloud service, exe.dev, provides VM pools with SSH and web auth designed for the agent era, ensuring secrets are not accessible to LLMs.

Performance & Benchmark

Attribution

Contribution of the three sectors within AGIX holdings today

Application: 0.53%

Hardware: 0.96%

Infrastructure: 1.93%

Top-performing stocks in AGIX holdings today

Twilio Inc. (19.36%)
Arm Holdings plc (15.73%)
Nebius Group N.V. (14.46%)
CoreWeave, Inc. (13.96%)
Datadog, Inc. (12.19%)
Cloudflare, Inc. (11.99%)
Zscaler, Inc. (11.44%)
Cadence Design Systems, Inc. (10.46%)
Samsara Inc. (10.17%)
Samsung Electronics Co., Ltd. (10.09%)
Oracle Corporation (9.91%)
Salesforce, Inc. (9.68%)
Snowflake Inc. (9.63%)
ServiceNow, Inc. (9.24%)
SAP SE (8.14%)
Nutanix, Inc. (7.3%)
Arista Networks Inc (7.03%)
NITTO BOSEKI CO LTD (6.93%)
Palo Alto Networks, Inc. (6.67%)
Micron Technology, Inc. (6.64%)
NVIDIA Corporation (6.26%)
Hut 8 Corp. (6.05%)
MediaTek Inc. (5.68%)
Tempus AI, Inc. (4.62%)
Shopify Inc. (4.56%)
Riot Platforms, Inc. (4.21%)
Broadcom Inc. (2.95%)
Palantir Technologies Inc. (2.63%)
Vertiv Holdings Co (2.43%)
Microsoft Corporation (2.28%)

Worst-performing stocks in AGIX holdings today

Amazon.com, Inc. (-3.47%)
NEXTERA ENERGY INC (-3.85%)
Texas Instruments Incorporated (-4.08%)
Tesla, Inc. (-4.57%)
Meta Platforms, Inc. (-5.07%)
Astera Labs, Inc. (-6.64%)
CIRCLE INTERNET GROUP INC (-7.11%)

Top five listed stocks contributing the most within AGIX holdings today

Arm Holdings plc (15.73%)
Nebius Group N.V. (14.46%)
NVIDIA Corporation (6.26%)
Snowflake Inc. (9.63%)
Datadog, Inc. (12.19%)

Observation

The strongest movers across peer ETF holdings today are listed below:

News

Funding & M&A

Anthropic files to go public

Artificial intelligence research lab Anthropic has confidentially filed its IPO application with the U.S. Securities and Exchange Commission, and after a recently completed $65 billion Series H funding round, its valuation has reached $965 billion. The round was co-led by multiple institutions including Altimeter Capital and Sequoia Capital, attracting a large number of strategic investors. Meanwhile, SpaceX is also preparing for an IPO with a target valuation of $2 trillion, while OpenAI has just completed a $122 billion funding round and plans an IPO. Anthropic's current annualized revenue is moving from $30 billion toward $47 billion, and it has introduced the Mythos model to the European Union's cybersecurity agency. This series of moves indicates that AI giants have entered IPO season, and the market will face a severe test.

SpaceX says it may issue ‘significant’ equity in ‘future transactions’

Space Exploration Technologies Corp. (SpaceX) warned investors in its first IPO filing amendment that it may issue a large amount of equity in future transactions. This statement was added to the risk factors section and is seen as preparation for a potential merger with Tesla. The company acquired Elon Musk's AI company xAI in 2025 and signed an option agreement with Cursor, allowing it to acquire that startup for $60 billion in stock after the IPO. SpaceX plans to raise $75 billion through the IPO, with $20 billion used to repay debts from xAI and X. Musk holds absolute voting power in SpaceX, but any merger would still require a vote by Tesla shareholders and regulatory review. This move could lead to significant equity dilution and affect investor returns.

Water access is now a risk factor in SpaceX’s IPO

SpaceX has added a risk warning regarding water resources for data center cooling in its IPO filing, emphasizing that water access is as important as core resources like electricity and processors. After incorporating Elon Musk's AI project xAI, the company notes that large-scale data center operations may require significant cooling water, and water shortages, drought conditions, or policy restrictions could limit cooling capacity, increase costs, and delay expansion plans. This revision may be related to SEC comment letters. Additionally, SpaceX plans to reserve up to 5% of IPO shares for employees and friends of executives, and warns that future large-scale issuance of new shares or a merger with Tesla could lead to shareholder dilution. This move highlights the growing resource constraints in AI infrastructure construction.

Anthropic confidentially files for an IPO amid rapid growth

Artificial intelligence company Anthropic has secretly submitted an IPO application, with an expected market value exceeding $96.5 billion after listing. The company's latest financial report shows annualized revenue has exceeded $47 billion, and it may have achieved its first profitable quarter, with the Claude Code product generating over $2.5 billion in annualized revenue. Anthropic plans to launch a Mythos-level LLM model and will collaborate with AWS, Broadcom, Google, and SpaceX to build large-scale infrastructure. SpaceX also submitted an IPO application the same week, and OpenAI is expected to go public within the year. This listing will unlock opportunities for institutional investors while providing infrastructure financing support for Anthropic.

Technical Breakthroughs

Five thoughts from Nvidia CEO Jensen Huang’s GTC Taipei 2026 keynote

Nvidia CEO Jensen Huang, in his keynote at GTC Taipei 2026, declared the official arrival of the 'Agent Era,' emphasizing that agentic AI has moved from experimentation to actual economic contribution. He noted that tokens have become directly monetizable profit units, and AI factories are becoming the world's largest infrastructure projects, with single-gigawatt investments reaching $80 billion to $100 billion. Huang unveiled the Vera Rubin multi-rack system and the Vera CPU, aiming to maximize token output per watt through a full-stack DSX solution, transforming Nvidia from a GPU company into an AI infrastructure company. The keynote stressed that AI will create more software development jobs rather than replace them, and that enterprises need to redesign system architectures to match agent workflows. Huang believes that computing power has directly converted into revenue, and architectural choices will determine future competitiveness.

Nvidia gives developers the tool to build secure, autonomous AI workers that scale.

At the Taipei GTC 2026 conference held concurrently with Computex, Nvidia officially released the latest version of Agent Toolkit, providing developers with tools to build long-running AI agents. The toolkit includes the NemoClaw framework, the Nemotron 3 Ultra 550-billion-parameter mixture-of-experts model, the OpenShell secure runtime environment, and multiple CUDA-X plugin skills. Through collaborations with Microsoft, Canonical, IBM Red Hat, and others, Nvidia has enhanced agent security and executability, supporting complex workflows, programming, research, and cybersecurity tasks. Early adopters such as Cadence, Siemens, CrowdStrike, and Palantir have applied it to chip verification, EDA workflows, vulnerability remediation, and offline autonomous systems. Nvidia CEO Jensen Huang stated that AI agents must operate securely and efficiently within enterprise systems to drive the AI revolution. The toolkit will be rolled out globally through platforms like Hugging Face.

Nvidia ramps up production of Vera Rubin, the foundation of the next generation of AI factories.

At the Taipei Computex conference, Nvidia announced that the Vera Rubin platform has entered mass production and will serve as the core infrastructure for agentic AI factories. The platform integrates multiple new technologies including Rubin GPU, Vera CPU, NVLink 6, and BlueField-4, delivering 10x agentic AI throughput. The new NVL72 liquid-cooled rack system can train mixture-of-experts models with a quarter fewer GPUs, reducing inference costs to one-tenth. Nvidia also introduced the Spectrum-X photonic network and the DOCA security framework, enabling full-stack confidential computing. The platform will begin shipping this fall, with over 350 suppliers including Dell, HPE, SuperMicro, and Lenovo participating. This launch marks the official move of the agentic AI era toward enterprise applications.

Intel touts over 130 edge design wins for Series 3 and launches the OpenVINO Physical AI framework.

Intel recently announced that its third-generation processor series has secured over 130 design wins for edge AI and edge computing, and launched a new open-source framework, OpenVINO Physical AI, aimed at bridging the gap between lab robot models and actual factory deployment. This processor series uses Intel 18A process technology, integrating CPU and AI acceleration to help robots transition from distributed architectures to a single SoC solution. The retail robot Ella has successfully migrated to the Intel platform, with three AI agents running in parallel on the same chip, significantly reducing system complexity and total cost of ownership. OpenVINO Physical AI provides a unified inference path from experimentation to mass production, supporting integration with open-source environments like LeRobot. Intel also released benchmarks showing its platform's performance on vision-language-action models is comparable to Nvidia Thor, but at half the system cost. A preview of the framework is now available on GitHub, with the official version planned for release in the second half of 2026.

DuckDuckGo makes its ‘no-AI’ search engine easier to access as its traffic booms

As an alternative search engine, DuckDuckGo is capitalizing on user backlash against AI search by launching a new browser extension. The extension allows users to set the AI-free search page noai.duckduckgo.com as their default search engine, avoiding AI-generated answers, chat prompts, and large amounts of AI images. The extension is now available on Chrome and Firefox, and the Privacy Essentials extension will later be updated to provide AI setting controls. Since Google announced its major AI-first search upgrade at its developer conference, DuckDuckGo's AI-free page traffic has grown nearly 30% week-over-week, with US app installations increasing 18.1% year-over-year, and single-day traffic on May 28 tripling. The company also continues to offer an AI chatbot and subscription services.

Nvidia gives developers the tools to build secure, autonomous AI workers that scale

During GTC Taipei 2026, Nvidia launched a new Agent Toolkit designed to help developers build long-running AI agents. The toolkit includes the NemoClaw framework, the Nemotron 3 Ultra 550 billion parameter model, and the OpenShell secure runtime environment, effectively addressing challenges in memory management, third-party tool invocation, and security control. OpenShell, developed in collaboration with Microsoft, Canonical, and IBM Red Hat, supports sensitive data masking and local hardware-first routing. Companies such as Cadence, Siemens, CrowdStrike, and Palantir have already applied the suite to automate chip verification, PCB design, and cybersecurity. Nvidia emphasized that for the AI revolution to truly take hold, agents must be seamlessly embedded into enterprise business systems. Nemotron 3 Ultra is scheduled for release on June 4, while NemoClaw and CUDA-X skills are already available, marking a key step in moving AI agents from concept to large-scale deployment.

This AI weather startup is out-forecasting government agencies

Founding company Windborne Systems today released the sixth version of its AI weather forecasting model, WeatherMesh 6, which surpasses both traditional and AI systems from the European Centre for Medium-Range Weather Forecasts (ECMWF) in accuracy for key variables such as surface temperature. The model leverages high-frequency data collected from the company's approximately 400 weather balloons worldwide, increasing forecast frequency to once per hour with a resolution of 3 kilometers. Compared to traditional physical models, WeatherMesh 6 uses a transformer architecture to directly merge data, maintaining high stability without relying on ECMWF initial conditions. The company has secured $25 million in funding at a valuation of $85 million, and its data is already used by NOAA, the U.S. Air Force, and the Navy. This technological advancement demonstrates the end-to-end advantages of AI in weather forecasting, from data collection to model training.

Fingerprint launches AI Assistant Detection to spot traffic from ChatGPT, Gemini, and Claude.

Device intelligence company FingerprintJS Inc. has released previews of two new products designed to identify traffic from AI assistants, addressing detection gaps that have emerged with increased AI access. AI Assistant Detection monitors traffic from AI assistants such as OpenAI's ChatGPT, Google's Gemini, and Anthropic's Claude in real time, and runs on the Automation Intelligence API, which identifies automated traffic without requiring client-side JavaScript. The new tools build on the Authorized AI Agent Detection released in February, extending coverage to content browsing and summarization AI assistants. With the emergence of browser-independent AI applications like Gemini Spark, traditional security tools based on browser signals are no longer adequate. The system performs IP range matching and DNS verification at the HTTP level, and also provides risk signals for proxies, VPNs, etc., helping enterprises distinguish genuine AI traffic from fake. This prevents malicious crawlers from impersonating AI user agents for data exfiltration, giving security teams more precise access control capabilities.

Nvidia chases the $200 billion CPU market with AI agent PCs from Microsoft, Dell, and HP

Nvidia launched the RTX Spark super chip at the Computex exhibition in Taipei, bringing a major technological breakthrough to the AI PC market. The chip delivers one petaflop of floating-point performance, can safely run AI agents such as OpenClaw and Hermes Agent, and supports local large language model inference. Multiple PC manufacturers including ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI will launch related products this fall, with Acer and Gigabyte following later. The chip integrates Nvidia CUDA software and supports AI features in over 1,000 games and applications, with software vendors like Adobe and Blender already signed on. CEO Jensen Huang proposed a new vision for a $200 billion CPU market, aiming to achieve a no-interface experience through voice interaction. Previously, ARM-based Windows devices failed due to poor performance, but the RTX Spark is powerful, and Microsoft calls its Surface Laptop Ultra the most powerful in history. Product pricing has not yet been announced, and it will compete with the Mac Mini in the high-end AI market.

Product Launches

Nvidia debuts the RTX Spark processor for Windows laptops and compact desktops

Nvidia launched the RTX Spark system-on-chip at the GTC Taipei event, based on the Blackwell architecture GPU and Grace CPU, offering 6,144 CUDA cores, 128GB of unified memory, and 1 PFLOPS of FP4 performance. The chip can support local LLMs with 120 billion parameters and a 1 million token context, suitable for AI agents, creative software, and gaming. In collaboration with MediaTek, it is adapted for Windows systems, and Microsoft will launch the Surface Laptop Ultra. Additionally, the OpenShell secure sandbox tool and DLSS 4.5 Ray Reconstruction technology were introduced, with vendors like Adobe adapting their AI features. This move represents a major restructuring of PC architecture in 40 years and will reshape the Windows AI experience.

With DGX Station for Windows, Nvidia squeezes a 1 trillion-parameter AI supercomputer into a deskside form factor.

Nvidia recently unveiled the world's first desktop AI supercomputer at the Taipei GTC conference—the DGX Station for Windows. Developed in deep collaboration with Microsoft, the system is built on the GB300 Grace Blackwell Ultra Desktop Superchip, delivering 20 petaflops of FP4 performance and 748GB of memory, capable of running frontier models with up to 1 trillion parameters locally or supporting hundreds of AI agents in parallel. Through Nvidia's OpenShell secure runtime and Windows security mechanisms, enterprise developers can build, deploy, and manage AI agents directly in the Windows environment without pushing initial workloads to cloud data centers. The system can also be configured with an RTX Pro 6000 Blackwell GPU to support physical AI simulation and visualization workflows, and multiple units can be networked via ConnectX-8 SuperNIC to form clusters. The product will be available in the fourth quarter of 2025, offered by vendors including Dell, HP, ASUS, MSI, and Super Micro.

Policy & Regulation

Hackers hijacked Instagram accounts by tricking Meta AI support chatbot into granting access

A recent security incident on the U.S. social platform Instagram involved account takeovers using Meta's own AI-powered support bots. Hackers, after spoofing their target's geographic location via VPN, asked the AI assistant to add a new email and obtain a verification code, ultimately resetting the password and taking over accounts. Compromised accounts included the former White House Instagram account and the U.S. Space Force Senior Enlisted Advisor account. Researcher Jane Wong also suffered a similar attack. Meta announced it has fixed the potential vulnerability but has not disclosed the number of affected users. This incident exposes the potential risks of AI customer service mechanisms in identity verification and reminds platforms to strengthen security strategies.

Florida sues OpenAI and Sam Altman in a first-of-its-kind lawsuit over violent incidents

Florida Attorney General James Uthmeier recently filed the first state-level lawsuit against OpenAI and its CEO Sam Altman, accusing the company of ignoring internal and external safety warnings in pursuit of an AI arms race and massive profits, leading to ChatGPT being used to assist shootings, induce suicide, and addict minors. The lawsuit also alleges that OpenAI collected children's data without parental supervision. Previously, the state had launched a criminal investigation into OpenAI regarding the shooting at Florida State University, and victims' families have filed civil claims for compensation. This lawsuit is the most stringent state-level AI accountability action to date and will influence global AI regulatory trends. OpenAI has just concluded its lawsuit with Elon Musk and now faces more civil lawsuits involving suicide and homicide.

Industry Partnerships

Merge launches Agent Handler for Employees as an IT gatekeeper for workplace AI agents

Capital provider Merge API Inc. has launched a new product, Agent Handler for Employees, solving the AI tool connectivity challenge for enterprise IT teams. The platform integrates with identity providers to automatically import employees and organizations, mapping them to approved AI tools and operations while enforcing data loss prevention and full log control. This effectively prevents employees from bringing in unauthorized external AI services, which is particularly valuable in regulated industries such as finance, healthcare, and government. Meanwhile, Perplexity AI has already used this service for enterprise solutions to build customer data connectors more quickly. Merge's technical background enables it to transition from static integration to dynamic model context protocol tool access, providing a new solution for secure and controllable agentic AI applications.

Mitel CX solves the sovereign cloud problem for communications

Sovereign cloud topics are shifting from theory to procurement reality. The traditional telecommunications industry has been slow to adopt AI due to its mission-critical nature, but low-barrier AI use cases like customer service are driving sovereign cloud to become a standard procurement requirement. European and American corporate boards emphasize data residency requirements, and regulators impose heavy fines for cross-border data governance failures. The Mitel CX platform treats sovereignty as an architectural principle rather than a deployment option, offering multiple levels of private, managed, trusted, and sovereign deployment, supporting private cloud, on-premises, and hybrid environments. It integrates generative AI virtual agents while maintaining data residency boundaries, making it suitable for highly regulated industries such as finance, healthcare, government, and hospitality. IDC data shows that half of enterprises have already adopted hybrid communication solutions, and 60% plan to replace platforms to meet compliance requirements. Mitel has deployed in the UK and multiple European countries, and has expanded into hospitality businesses. Its multi-vendor management capabilities and product-level governance architecture give it a competitive advantage over vendors like AWS and Microsoft. Sovereignty has become a core competitive factor rather than a mere technical issue.

融资收购

On theCUBE Pod: Dell’s AI hardware business pays dividends; Snowflake and Anthropic fare well at the stocks

This week, Dell Technologies, powered by AI servers, saw revenue far exceed expectations with 88% growth, and its stock price rose over 33% in a single week, becoming a benchmark for hardware stocks. Its pre-configured racks and high GPU utilization products are highly favored by Neocloud and enterprise customers. Meanwhile, NetApp's stock also rose over 30%. Snowflake's stock rose 36% ahead of its upcoming Summit, and it committed to increasing AWS computing purchases. Anthropic just completed $65 million in funding, surpassing OpenAI in amount. Analysts point out that AI infrastructure investment still faces obstacles such as low utilization rates and cooling method transitions, and future growth may experience volatility. The performance and funding trends of data giants reflect that the AI hardware boom is still encountering new opportunities and adjustments.

Alphabet plans to raise $80 billion to pay for its AI buildout

Google parent company Alphabet announced on Monday that it will raise $8 billion through a stock offering to support large-scale AI infrastructure construction. The plan includes selling $10 billion in stock to Berkshire Hathaway. The company stated that current enterprise and consumer demand for AI services has exceeded its supply capacity, and it is rapidly expanding infrastructure to meet market growth. This move will maintain a healthy balance sheet rather than relying entirely on debt financing. Similar to other tech giants, Alphabet plans to invest $18-19 billion in computer capital expenditures in 2026, and global tech companies' AI-related capital expenditures are expected to reach $700 billion. This financing will help strengthen its competitive advantage in the AI field.

Alphabet unveils a plan to sell $80 billion in shares to fund its ongoing AI infrastructure buildout

Alphabet Inc. plans to raise $80 billion in equity capital through stock offerings to support its artificial intelligence infrastructure construction. Berkshire Hathaway has agreed to invest $10 billion through a private placement. Alphabet stated that the funds will be used for data centers and world-class AI computing power investments to meet unprecedented demand for AI services from enterprises and consumers. Alphabet is accelerating its efforts to catch up with competitors such as Amazon Web Services, Microsoft, and Meta. The four major cloud giants are expected to spend a combined over $700 billion on AI capital expenditures this year. Alphabet has raised its full-year capital expenditure forecast to between $180 billion and $190 billion. CEO Sundar Pichai said computing capacity is his biggest concern. Alphabet has recently raised over $66 billion through multiple rounds of bond issuances and will also issue $30 billion in underwritten stocks, $15 billion in mandatory convertible preferred shares, and sell $40 billion in Class A and Class C stocks in the third quarter through an ATM program. Goldman Sachs, JPMorgan Chase, and Morgan Stanley are acting as underwriters.

RSS Subscription

Do not be fooled

Charles Mackay's 1841 book "Extraordinary Popular Delusions and the Madness of Crowds" explores how easily people become captivated by certain ideas, emphasizing that the timing and demand for an idea matter more than its quality. In modern society, the internet accelerates this phenomenon, as people often accept and spread popular views without deep thought, leading to rigid thinking and a lack of self-awareness. To avoid this, it is recommended that individuals reflect on their own positions, engage with those who hold different views, and deeply understand opposing arguments.

The web is changing, and we are not going back

As Google's search engine gradually shifts toward natural language processing, users no longer need to use machine language to search; Google can understand natural language and provide answers. However, this change has also made it more difficult to obtain information, as users often cannot verify the accuracy of answers because most people no longer click on source links. Nevertheless, other search engines such as DuckDuckGo and Kagi are still available, allowing users to adjust according to their needs without relying entirely on Google.

Hackers Used Meta’s AI Support Bot to Seize Instagram Accounts

The Instagram accounts of the Obama White House and the Chief Master Sergeant of the U.S. Space Force were recently hacked. The hackers exploited a vulnerability in Meta's AI support assistant to reset account passwords, successfully tampering with these accounts and posting pro-Iranian content. The hackers claimed to have disguised themselves as target users via VPN, used the AI assistant to link the accounts to new email addresses, and thereby obtained verification codes to reset passwords, hijacking multiple short accounts valued at over $500,000. Meta has urgently fixed the issue and stated that no database breach occurred. Security experts warn that as more platforms allow AI chatbots to handle sensitive account recovery requests, more such attacks may emerge in the future.

Stealth Startup Spy #344

Former Apple Siri and Roblox machine learning scientists have entered stealth mode, a former Ogilvy partner is building an AI platform supporting the full lifecycle of private equity, and a former Lyft hardware director has exited stealth with a $12 million seed round, aiming to let engineers test hardware changes in seconds.

It’s not just Taylor series

Discussion continues regarding the approximation formula exp(−x²) ≈ (1 + cos(sin(x) + x))/2, with some believing it can be explained via Taylor series. Although the two first differ at the x^6 term, leading to a good approximation, the approximation is valid over the interval [−4, 4], where the Taylor series error is far larger than that of the cosine approximation. At x = 4, the error of the cosine approximation is 0.002579, while the Taylor approximation error reaches 113, a difference of five orders of magnitude.

Subscribe by email

Since the blog was founded in 2008, readers have experienced many changes in how they subscribe via email. Over the past two years, I have been using Substack to send notifications of new blog posts.

Intel 8088s and non-Intel non-clones

The Intel 8088 CPU made its debut on June 1, 1978, becoming the core processor for the IBM PC and its clones, with sales reaching approximately 60 million units. In addition to Intel, at least ten other companies obtained licenses to produce the 8088, including AMD, Fujitsu, and NEC. Although NEC developed an improved version, the V20, and won a legal battle, market demand for the 8088 series had significantly declined by the time the legal issues were resolved.

Pluralistic: Molly Crabapple's 'Here Where We Live Is Our Country' (01 Jun 2026)

Molly Crabapple's "Where We Live Is Our Country" explores the history of the Jewish Labor Bund, a socialist organization that dominated Jewish political identity in the late 19th and early 20th centuries. Although little known today, the Jewish Labor Bund played a significant role in the Russian Revolution and the anti-fascist struggle, highlighting the intersection of universal liberation struggles with identity politics. The book reveals a "mirror world" of Jewish identity, contrasting the different ideologies of the Jewish Labor Bund and Zionism.

1,000 Data Breaches Later, the Disclosure Lag is Worse Than Ever

Today, I recorded the 1000th data breach in "Have I Been Pwned." Although privacy regulations such as GDPR and CCPA have been introduced since I launched HIBP, the problem of delayed disclosure of data breaches has become more serious. For example, Carnival Cruise Line took 43 days to notify affected users after suffering a data breach, during which time the leaked data spread widely online. This delay may be related to the subsequent class-action lawsuits, as companies often prioritize legal liability over customer protection when handling data breaches.

Weekly Update 506

The current ShinyHunters breach has attracted widespread attention. Although there is clear criminal activity behind it, the organization's response (or lack thereof) to the victims is equally noteworthy.

Weekend trivia: your process' memory is a file

In Unix systems, although the design philosophy that 'everything is a file' is not entirely true, the /proc/<pid>/mem file provides a unique way to access process memory. Through specific system calls, the memory of a target program can be read or modified in real time. This method was first discovered 25 years ago and has been updated for modern 64-bit systems. Recently, the author rewrote a program called memfetch, which can capture memory snapshots of any process and generate binary files for analysis.

Weird projects I shipped with AI

Over the past year, I have used artificial intelligence to build a series of personal projects, including a daily game website based on the classic game SkiFree called skifreedle.com, Autodeck which automatically generates Anki cards, and an AI-generated endless wiki that has produced over 280,000 pages of content. The implementation of these projects was made possible with the help of AI, allowing me to accomplish tasks that would have been difficult to complete without AI.

Checking assembly with Z3

A new contributor to ZJIT, dak2, submitted a fix PR that resolves an overflow error when dividing FIXNUM_MIN by -1 in fixnum division. The fix handles this special case with a branchless test, ensuring the result is correctly promoted to bignum instead of fixnum. After verification using the Z3 SMT solver, the correctness of the new code was confirmed.

Joining the IndieWeb - #1: Microformats

This article is the first part of a series, documenting how the author made their website more interoperable with IndieWeb tools and standards. The author emphasizes the importance of interoperability in the digital world and introduces microformats, an open standard that describes content through simple HTML tags, aimed at enhancing information sharing and connections between websites. By using these standards, the author hopes to support the decentralization of the digital world and promote a more social internet environment.

Hackers simply asked Meta AI to give them access to high-profile Instagram accounts. It worked.

A hacker successfully requested to associate a high-profile Instagram account with a new email address by chatting with Meta's AI support bot, revealing a security vulnerability in Meta's support system. This incident highlights the risks of directly connecting support bots with account recovery processes.

'We Are Living in Pinocchio's World'

Writer and investor Om Malik has a deep understanding of Italian author Carlo Collodi's 'The Adventures of Pinocchio', believing the work reveals a society full of deception. Although many only remember the lies and the long nose from the story, Collodi's tale actually explores the complexity of missing authority and responsibility, reflecting the half-truths of modern society. Malik has a special fondness for Montblanc's limited edition Collodi pen, considering it not only beautiful but also rich in story, symbolizing Collodi's profound message.

Amazon Made AI Podcasts for Products

Amazon has launched a new feature that uses artificial intelligence to generate podcast-like audio clips, where two AI 'hosts' discuss the pros and cons and user reviews of specific products. The feature is based not only on product descriptions and customer reviews but also integrates information from other online sources, aiming to help users learn about products more easily. Although the effect is limited on some products, this innovation may appeal to consumers who prefer learning through listening.

The Infosec Phrasebook

The security field has a unique set of terminology, containing many militarized expressions and technical terms, such as 'defense in depth' referring to coding, 'zero trust' referring to authentication, and 'attack surface' referring to code. Additionally, the word 'cyber' is widely used, forming new concepts like 'cyber risk' and 'cyber resilience', reflecting the importance of cybersecurity in modern society.

May 2026 newsletter

I just released the May 2026 patron-exclusive monthly newsletter, covering topics such as Al's fee increase, Anthropic's good performance, disappointment with model releases, conference and podcast information, as well as the launch of Datasette Agent and progress on Datasette. Patrons can get early access to the newsletter by paying $10 per month.

I'm not sure we'd use AppArmor much even if we could

Due to attacks from high-traffic crawlers, you may encounter a blocking page when visiting my blog (Wandering Thoughts) or its related CSpace. This is because your browser version is considered too old, especially Chrome. Please ensure you are using the latest version of your browser and contact me for further assistance. It is recommended to use archive.org for page archiving, as other archiving services may misuse old user agents, leading to access issues.

The Talk Show Live From WWDC 2026: Tuesday June 9

On June 9, 2026 at 7 PM, the annual live "Talk Show" will be held at the California Theatre in San Jose, with tickets priced at $45. Special guests will attend the event, and the audience can also interact with other readers and listeners before the show begins.

exe.dev

exe.dev is a cloud service designed for the agent era, providing a pool of virtual machines with SSH, root, and web authentication by default, ensuring that secrets injected at the network edge are not obtained by large language models. Users can easily share web servers like sharing a Google Doc, and pay based on underlying resources rather than per virtual machine.

Take Two

Apple AI employee Kelsey Peterson, who was responsible for the unreleased 2024 Siri overhaul, has now joined OpenAI, so next month's WWDC will feature a new host to introduce the next-generation Siri.

datasette 1.0a32

datasette 1.0a32 released a small fix that addresses a bug with INSERT ... RETURNING queries via the new /db/-/execute-write endpoint, as well as some base_url issues encountered when experimenting with Service Workers.

The art and engineering of Silpheed

The 1990s was a remarkable decade for video game console development. The advent of CD-ROM drives offered 640 MiB of storage, but their access times and bandwidth lagged far behind cartridges. Sega's Mega-CD launched nearly 200 games for its Genesis console, among which "Silpheed" was highly praised for its outstanding art design and animation effects, despite many FMV games suffering from the technical limitations of the time. Developer Game Arts made smart trade-offs by working within the system's constraints, successfully achieving the game's visual effects under limited color and technical conditions.