🎙️ AI 日報 Podcast，自主 AI 駭客、Anthropic 營收超車、開源 GLM-5.2 登頂

Opening

Hello everyone, and welcome to "Mark's Tech Insights"! Today, we're going to talk about some jaw-dropping, blockbuster AI news—from the first-ever fully autonomous offensive AI ransomware to Anthropic surprisingly overtaking OpenAI in revenue, and a historic milestone for the open-source community. We have a massive amount of information packed into today's episode, so get ready for an exciting ride. Let's dive right in!

Today's Top Stories

1. A Historic First! AI Agent Autonomously Executes End-to-End Ransomware Attack "JADEPUFFER"

Source: The Hacker News (https://thehackernews.com/2026/07/ai-agent-exploits-langflow-rce-to.html）
Summary: Security firm Sysdig has discovered a new threat named JADEPUFFER, a fully LLM-driven AI Agent ransomware. Exploiting a security vulnerability in Langflow, it autonomously completed credential theft, lateral movement, encrypted 1,342 database records, and even self-debugged login failures along the way—all in just 31 seconds without any human intervention.
Taiwan Perspective: Many Taiwanese enterprises and startups are currently actively adopting tools like Langflow to develop AI applications. This incident sounds an alarm; development teams absolutely cannot take security lightly when pursuing AI implementation.
Discussion Points:
- With AI possessing "self-debugging" and "autonomous decision-making" offensive capabilities, how much longer can traditional security defense networks hold up?
- Will we need "AI vs. AI" automated defense mechanisms in the future to intercept these types of sub-minute attacks in real-time?
Script Suggestion: Everyone really needs to be on high alert. We used to say AI is at most a coding assistant, but who would have thought it can now act as a hacker on its own! The most terrifying part of JADEPUFFER is that when it encountered a login failure, it actually debugged itself and got everything done in 31 seconds. Humans simply cannot react to that speed. Many of our teams in Taiwan love using Langflow to build workflows right now. Please go back and immediately check your system updates; don't let your AI assistant become an insider threat.

2. The Golden Crossover! Anthropic's Annualized Revenue Hits $47 Billion, Overtaking OpenAI

Source: AI Weekly / Fortune (https://aiweekly.co/ai-news-today/anthropic-news）
Summary: Confirmed by Fortune magazine, Anthropic's momentum is incredibly staggering. Its projected annualized revenue for 2026 has reached $47 b i l l i o n, w i t h p r o f i t a b i l i t y i n s i g h t . I n c o n t r a s t, f o r m e r l e a d e r O p e n A I^{'} s l a t e s t p r o j e c t e d a n n u a l i z e d r e v e n u e s i t s a r o u n d$ 25 billion to $33 billion, meaning Anthropic's revenue scale is now nearly double that of OpenAI.
Taiwan Perspective: This indicates that enterprise clients highly accept the Claude series, which focuses on "safety and Constitutional AI." When major Taiwanese financial holdings and high-tech manufacturing industries choose enterprise-grade AI, security and compliance are often their top priorities.
Discussion Points:
- Is the key to Anthropic's successful takeover because enterprise customers trust their security more, or has the product capability truly widened the gap?
- How should OpenAI respond to this revenue lag? Will it accelerate the development of GPT-6?
Script Suggestion: This is truly a historic moment. Who would have thought that the Anthropic team, who originally spun out of OpenAI, would overtake their former employer in revenue today—and by nearly double the margin! This proves that business is business; enterprise clients don't want a celebrity hyped up on social media every day, but rather a stable, secure AI service that protects business secrets. When Taiwanese enterprises evaluate AI adoption, they can look to this trend: secure and compliant solutions are the ones that laugh last in the market.

3. Google Hits the Brakes! Gemini 3.5 Pro Delayed for a "Ground-Up" Architecture Rebuild

Source: BigGo Finance (https://finance.biggo.com/news/6f0c6bb2-795f-4c57-9d09-6db691d7638a）
Summary: Google DeepMind announced that it is postponing the release of Gemini 3.5 Pro to July 17, 2026. The reason is shocking: the official team decided to completely scrap the original 2.5 Pro architecture and perform a thorough architectural rebuild. The goal is to drastically improve mathematical reasoning, SVG scene generation, and image quality, while providing a massive 2-million-token context window.
Taiwan Perspective: Although the delay is disappointing, for Taiwanese engineers who need to handle long documents and complex code, the rebuilt 2-million-token window and stronger reasoning capabilities are well worth the wait.
Discussion Points:
- Google choosing to delay and "start from scratch" reflects just how intense the current competition among large models has become.
- Will the pricing model after the rebuild be competitive in the market?
Script Suggestion: Google is really going all out this time, actually deciding to scrap the original architecture and rewrite it at the last minute! You can tell they are under immense pressure facing the steady advance of Anthropic and OpenAI; they absolutely cannot afford to put out another mediocre upgrade. Although we have to wait until July 17th, if the rebuild can elevate mathematical reasoning and image quality to a whole new level, combined with that massive 2-million context window, it will definitely be a productivity powerhouse for the Taiwanese tech industry, which needs to process large volumes of complex specification sheets.

4. Global AI Funding Surpasses $510 Billion in H1 2026, Setting a Historic High

Source: Crunchbase News (https://news.crunchbase.com/venture/global-startup-exits-ipo-ma-soar-ai-q2-h1-2026/）
Summary: In the first half of 2026, global venture capital investment in the AI sector reached a staggering $510 b i l l i o n, d i r e c t l y s u r p a s s i n g t h e$ 440 billion raised in the entirety of 2025. Among them, the two giants OpenAI and Anthropic accounted for $217 billion, representing 43% of the total funding. Nearly 40 new AI unicorns were also born during this period.
Taiwan Perspective: Although massive amounts of capital are concentrated on US giants, this wave of hot money has also spilled over into the application layer. Taiwanese AI startups should seize this wave of funding and actively seek opportunities to expand internationally.
Discussion Points:
- Will the high concentration of capital in a few giants lead to severe monopoly issues in the AI market?
- With $510 billion burned in just half a year, how much of this is a bubble, or has AI's commercial monetization capability already grown to support this valuation?
Script Suggestion: Listen to this number, everyone. In just the first six months of this year, the money poured into AI globally reached $510 billion, which is more than the entire last year! However, if you look closely, you'll find that nearly half of the money was sucked up by the two super-behemoths, OpenAI and Anthropic. This is a revelation for our startup community in Taiwan: we don't need to compete with others on foundational large models. Instead, we should focus on how to utilize the technology of these giants to build the most practical vertical application layers. That is where the golden opportunity lies for Taiwanese companies.

5. The Pride of Open Source! GLM-5.2 Mixture-of-Experts Model Claims #1 in Global Open-Source Benchmarks

Source: LLM Stats (https://llm-stats.com/llm-updates）
Summary: GLM-5.2, an open-source Mixture-of-Experts (MoE) model with 744 billion parameters, successfully claimed the top spot on the open-source weight model performance leaderboard in the latest evaluations. Released under the MIT license, the model actually activates about 40 billion parameters per token. This marks the first time in history that a fully permissively licensed open-source model has topped an authoritative third-party leaderboard.
Taiwan Perspective: The MIT license means that Taiwanese enterprises can deploy this model locally completely free of charge and without commercial restrictions, allowing for Traditional Chinese fine-tuning and private deployment, which significantly lowers development costs.
Discussion Points:
- Does GLM-5.2 taking the top spot mean that the technical gap between open-source and closed-source models has been completely closed?
- How will the MIT license change the ecosystem for SMEs and academia developing customized AI?
Script Suggestion: For those of you who love playing with open-source models, it's time to celebrate! GLM-5.2 not only took first place in open-source performance, but the most generous part is that it was actually released under the MIT license. What does this mean? It means you can modify it however you want, use it however you want, and even use it commercially to make money without paying any licensing fees. For many of our enterprises in Taiwan that value data privacy, don't want to upload data to the cloud, and want to do on-premise fine-tuning themselves, this is absolutely the best gift of the year. You can go ahead and download it to test it out right away.

Closing

Alright, every single one of these five stories today is a heavyweight. From breakthroughs in local open source to the revenue shakeup of cloud giants, and even the new challenges of cybersecurity defense, they all indicate that the pace of change in the AI era will only get faster. I hope today's content helped you grasp the latest tech pulse. If you like our show, don't forget to subscribe and share it with friends around you who are interested in AI. We'll see you next time. Bye-bye!