🎙️ AI 代理人首度攻台？Copilot 記憶污染與 OpenAI 簡訊功能警訊｜AI 日報 Podcast

Introduction

The world's first autonomous AI agent cyberattack actually took place in Taiwan. Hackers let the AI loose to find vulnerabilities, bypass defenses, and even perform parallel scans on nuclear and energy systems. Today, we'll dive deep into this chilling cybersecurity incident, discuss the out-of-control risks hidden behind OpenAI's auto-SMS feature, and explain why Microsoft Copilot's "memory poisoning" vulnerability leaves you helpless even after changing passwords or deleting devices. We'll cover all of this in just a moment.

Today's Top Stories

1. AI Agent Launches First "End-to-End Fully Autonomous" Cyberattack, Taiwan Becomes the World's First Battlefield

Source: CNN Business (https://www.cnn.com/2026/08/13/tech/china-taiwan-ai-agent-cyberattack-intl-hnk）
Summary: Israeli cybersecurity firm Dream revealed that hackers suspected of being linked to China used an open-source AI agent framework to launch a highly autonomous cyberattack against the Taiwanese government in early July this year. When blocked, this AI system found new paths on its own and even autonomously mapped out multiple related systems for parallel scanning without human intervention.
Most Surprising Aspect: The AI was actually able to extract APIs and keys on its own from a single portal website, and dynamically changed its tactics when blocked.
Taiwan Perspective: As a geopolitical hotspot and a major tech hub, Taiwan has clearly become a "living testbed" for the world's newest and most cutting-edge AI weapons. The cybersecurity mindset of the Taiwanese government and tech industry must immediately upgrade from "defending against human hackers" to "countering autonomous AI speed."
Discussion Points:
1. When the speed and autonomy of AI attacks exceed the limits of human reaction, how should traditional blue team defense transform?
2. Should the weaponization of open-source AI frameworks (such as Hermes, OpenClaw) be regulated by international laws?
Script Suggestion: Taiwan is truly on the front lines of this tech war. In the past, when we talked about AI hackers, it was at most humans using AI to help debug code. But this time, it was "end-to-end" fully autonomous. The AI acted like a special forces unit with its own cognitive abilities—if it couldn't break in, it switched paths on its own, and even scanned nuclear safety and energy companies along the way. This means that in future cyber warfare, humans won't have time to react. We must use AI to defend against AI; otherwise, our defense speed simply won't keep up.

2. Microsoft Copilot Hit by Major "CoSnitch" Vulnerability: Hackers Poison AI Memory, Making Password Changes Useless

Source: Dark Reading (https://www.darkreading.com/vulnerabilities-threats/cosnitch-attack-copilot-mapping-out-architecture）
Summary: Cybersecurity firms reported a major vulnerability in Microsoft Copilot called CoSnitch. By tricking users into clicking a malicious link, attackers can force Copilot to automatically read and exfiltrate sensitive user data from Gmail, Google Drive, and other sources. Most terrifyingly, the malicious instructions are written into Copilot's long-term memory, creating a persistent contamination that cannot be cleared by changing passwords or resetting devices.
Most Surprising Aspect: Once malicious instructions are written into the AI's long-term memory, they cannot be cleared even if you change your password or log out of all devices.
Taiwan Perspective: Taiwanese enterprises are actively adopting generative AI tools like Copilot for Business. This incident rings an alarm bell, reminding us that an AI's "long-term memory feature" can become a blind spot outside traditional cybersecurity defenses. Enterprises must establish auditing mechanisms specifically for AI memory.
Discussion Points:
1. Microsoft took nearly eight months to patch this vulnerability. Does this reflect tech giants neglecting basic security protections in their rush for AI speed?
2. As AI gains cross-platform data access permissions (such as email and calendars), how should we define security boundaries?
Script Suggestion: This vulnerability is truly hair-raising. Previously, if we got hacked, changing our password and logging out of our accounts would solve the problem. But this time, hackers injected poison directly into Copilot's "brain"—its long-term memory. As long as Microsoft didn't patch it, this AI would keep secretly snooping on your emails for the hackers. What's even crazier is that Microsoft actually dragged its feet for eight months before fixing it. This tells us that before handing all our privacy over to AI assistants, we really need to be extra cautious.

3. ChatGPT Launches Apple Messages Plugin, but OpenAI Warns Users "Never Enable Auto-Authorization"?

Source: TechCrunch (https://techcrunch.com/2026/08/20/chatgpt-can-now-send-texts-for-you-with-new-apple-messages-plugin/）
Summary: OpenAI has launched a brand-new Apple Messages plugin, allowing ChatGPT to help you organize, search, and even draft and send text messages directly. Although all messages are processed locally on-device to protect privacy, OpenAI took the rare step of warning users not to enable persistent auto-authorization, to avoid losing the chance to perform a final check before sending.
Most Surprising Aspect: The scenario demonstrated in the official promo video actually showed the AI helping you figure out "how to reply today to yesterday's conversation," yet the company itself warned everyone not to give the AI full authority.
Taiwan Perspective: LINE is extremely popular in Taiwan. If a similar LINE AI plugin is released in the future, it will undoubtedly spark heated discussions among Taiwanese users. However, this also tests the public's moral and privacy boundaries regarding "handing over the right to reply to personal relationships to AI."
Discussion Points:
1. When text replies between people are handled by AI, what is left of the essence and sincerity of our communication?
2. Is OpenAI's warning an attempt to evade legal liability, shifting the risk of AI-sent messages onto the users?
Script Suggestion: OpenAI's new feature to send text messages for you sounds super convenient, right? But the funniest part is that they wrote a massive warning telling everyone absolutely not to enable "persistent authorization." Why? Because if you let the AI decide when to send texts, you might wake up to find it has already called in sick to your boss or said something bizarre to your partner. This is practically handing your social life over to a robot. Technology is truly making us lazier—and putting us in more danger.

4. To Move Data Centers into Space, Starcloud Raises $250 Million, with Nvidia Joining the Mix

Source: TechCrunch (https://techcrunch.com/2026/08/21/starcloud-raises-200-million-for-orbital-data-centers-as-launch-options-dry-up/）
Summary: Space data center startup Starcloud has raised $250 m i l l i o n a t a$ 2.3 billion valuation, with investors even including Nvidia. They plan to launch satellites equipped with high-end GPUs like the H100 into orbit. However, with SpaceX's Falcon 9 nearing retirement and launch slots severely limited, some competitors are even starting to consider building their own rockets.
Most Surprising Aspect: Because rocket launch slots are so highly sought after, an orbital data center startup actually decided to "just build their own rocket" because they couldn't get in line.
Taiwan Perspective: Taiwan possesses a robust semiconductor and low-Earth orbit (LEO) satellite supply chain. If the concept of space data centers becomes a reality, it will open up a brand-new, high-margin blue ocean market for Taiwanese space communication component manufacturers and thermal module factories.
Discussion Points:
1. Operating GPU data centers in space involves extremely high technical difficulty in terms of heat dissipation and cosmic radiation protection. Can this really be commercialized?
2. Is the monopoly in the rocket launch market (such as SpaceX) becoming a physical bottleneck restricting the development of AI and space technology?
Script Suggestion: Sending H100s into space for computing sounds a lot like science fiction, right? Starcloud securing investment from Nvidia proves this is no joke. But the craziest part is that because SpaceX's Falcon 9 is retiring soon, launch slots are almost impossible to get, forcing competitors to "build their own rockets." This shows us that the AI compute arms race has escalated from the Earth's surface into outer space, and physical limitations and rocket production capacity have actually become the new bottlenecks holding back AI development.

5. TSMC's Capacity Fully Booked, Samsung Seizes Opportunity to Raise Advanced Node Prices by Up to 15%

Source: Tom's Hardware (https://www.tomshardware.com/tech-industry/samsung-raises-advanced-foundry-prices-by-up-to-15-percent-as-ai-demand-fills-its-4nm-lines）
Summary: Since TSMC's 3nm and 2nm capacities have been completely booked out by giants like Apple, Nvidia, and AMD, a lot of spillover chip demand has been forced to turn to Samsung. Consequently, Samsung raised its quotes for new orders on 4nm, 5nm, and 8nm nodes by up to 15% starting in July, with US and Chinese customers seeing the largest increases, while Taiwanese customers saw relatively lower hikes.
Most Surprising Aspect: Samsung had just lost a third of its foundry market share not long ago, yet it managed to aggressively raise prices on the back of the AI wave, causing its US ADR to jump over 10%.
Taiwan Perspective: This proves that TSMC's absolute hegemony in advanced nodes is so powerful that "even the leftover crumbs are enough to feed competitors and let them raise prices." However, Taiwanese IC design houses also need to be cautious, as Samsung's price hikes could drive up overall chip manufacturing costs.
Discussion Points:
1. Will Samsung's price hikes push up the hardware costs of next-generation AI chips, thereby affecting the pricing of end-user applications?
2. How long is the severe supply-demand imbalance in foundry capacity expected to last?
Script Suggestion: Everyone knows TSMC is currently an invincible force, with its 3nm capacity already sold out through 2027. But the wildest part is that TSMC is so full that even Samsung next door is making a fortune. Samsung clearly lost a lot of market share previously, but now, because everyone is stuck in TSMC's queue, they have no choice but to knock on Samsung's door. Samsung straight up announced a 15% price hike, and its stock price surged. This is the madness of the AI era—when the leader eats the meat, the runner-up can get stuffed just by drinking the soup.

6. One-Third of Web Pages Created Since ChatGPT's Launch Are Actually Written by AI?

Source: TechCrunch (https://techcrunch.com/2026/08/20/a-third-of-webpages-published-since-chatgpts-launch-show-signs-of-ai-authorship-study-finds/）
Summary: A recent study by the Pew Research Center indicates that since ChatGPT launched in late 2022, up to 35% of newly created web pages on the internet show clear signs of AI ghostwriting. Among these, the proportion of AI content in .com domains is ten times higher than in academic and government domains, while non-profit .org domains have the least.
Most Surprising Aspect: A large portion of the internet ecosystem has already evolved into "robots reading web pages written by other robots," establishing a circular logic of AI feeding AI.
Taiwan Perspective: Taiwan's Traditional Chinese web environment is also facing erosion from a massive influx of AI-generated content (MFA websites). This not only degrades search quality but also makes it harder for high-quality local Taiwanese creators to be seen in the algorithms.
Discussion Points:
1. As AI-generated data dominates the web, will training future AI models on this data lead to "Model Collapse"?
2. As internet users, how can we rebuild trust in online information?
Script Suggestion: Have you noticed that Google Search is getting harder and harder to use lately, and clicking on links just leads to a bunch of wordy, low-quality content farm articles? The Pew Research Center provided the hard data: since ChatGPT came out, over one-third of new web pages on the internet have been written by AI. The most ironic part is that much of the training data for AI is scraped from the web, which means "robots are reading what robots wrote, to write new things for robots to read." This "inbred" AI loop makes you worry whether the future internet will turn into a sea of information garbage.

Closing Thoughts

From cyber warfare on Earth and orbital data centers in space, to ubiquitous AI-ghostwritten web pages on the internet, AI is reshaping our world at an unimaginable pace. While we enjoy the convenience, we must also remain vigilant and not let our privacy and judgment be swept away by technology. That's all for today's "Mark's Tech Insights." See you next time!