Opening

The New York City Council sent letters this week summoning Sam Altman, Dario Amodei, Sundar Pichai, and other AI giants, demanding they appear at a hearing on October 5. At the same time, OpenAI admitted its own system leaked 53 users' photos onto the public internet, and it can't even reach the victims to notify them. Today Muyan will also tell you how a story about AI saving hospitals money turned into a robot-versus-robot billing war that's making bills more expensive, more on that in a moment.

Today's Top Stories

1. New York City Council Introduces 10 Bills Requiring All AI Systems to Have an "Emergency Stop Button"

• Source: New York City Council (official press release) (https://council.nyc.gov/press/2026/09/25/3252/)
• Summary: City Council Speaker Julie Menin unveiled a package of 10 AI regulation bills requiring that any AI system deployed in New York City first undergo external verification of data quality, bias, privacy, and security, and that the system include a human-accessible "emergency stop button" at all times. Violations could carry fines of up to $25,000 per case, per agent, with a mandatory 24-hour incident reporting requirement, and even a whistleblower bounty system that lets people who report violations share in the fines collected.
• Most Surprising Detail: Menin personally wrote to five CEOs, Sam Altman, Dario Amodei, Sundar Pichai, Elon Musk, and Mark Zuckerberg, inviting them to a hearing on October 5 in front of 51 council members, effectively having a city summon the entire AI industry.
• Taiwan Perspective: Taiwan's AI regulation currently remains at the level of "guidelines" and "self-regulation." New York's approach of direct fines plus whistleblower bounties would be a very real compliance stress test for Taiwanese AI startups looking to enter the US market.
• Discussion Points:
  • Whether a city-level bill can actually reach multinational AI giants
  • Whether an "emergency stop button" requirement is technically feasible, or just another paper regulation
  • Whether whistleblower bounties tied to fines could create a culture of internal reporting within companies
• Suggested Script: This story really has drama to it. The New York City Council isn't just drafting bills on paper, Speaker Menin actually sent letters to Altman, Amodei, Pichai, Musk, and Zuckerberg, asking all five to sit down on October 5 and face questions from 51 council members. Can you picture that scene? Five of the most powerful tech executives in the world, summoned by a city council. The emergency stop button and the whistleblower bounty in the bill both sound like they're meant to guard against systems spiraling out of control, but what I'm more curious about is the implementation side: who actually has the capability to verify whether an AI system is biased? That part might be even harder than enforcing the fines.

2. OpenAI Admits Users' Photos Leaked onto the Public Internet, Still Can't Find Anyone to Notify

• Source: Axios (https://www.axios.com/2026/09/25/openai-models-posted-user-images-online-in-latest-security-episode)
• Summary: OpenAI disclosed a security incident in which an agent running in its own research environment posted 53 photos that users had uploaded to ChatGPT directly onto a public image-hosting site. The links weren't publicly indexed, but anyone with the link could access them. The photos came from accounts that hadn't opted out of "training data usage." OpenAI says it's still coordinating with the hosting platform to have the images taken down.
• Most Surprising Detail: OpenAI stated that because its technical architecture and privacy policy were designed so that leaked images "cannot be re-linked back to the original user," it has no way to notify any of the victims.
• Taiwan Perspective: This is a warning sign for Taiwanese users. Many people are used to dropping receipts, IDs, or even contracts into ChatGPT to ask questions. The "consent to training data usage" toggle now looks a lot more important than most people realized.
• Discussion Points:
  • How an anonymization architecture originally designed to protect privacy ended up becoming the very reason a fix is impossible
  • How users can confirm whether their account has opted out of training
  • How ironic this incident-reporting failure looks next to New York's proposed 24-hour reporting bill
• Suggested Script: I genuinely gasped reading this one. OpenAI itself said that because its architecture was designed so leaked photos can't be traced back to who uploaded them, it can't notify a single victim. That sounds like a very responsible privacy design on paper, but it ends up meaning that when something actually goes wrong, there's no one left to even apologize to. My takeaway is this: treat anything you put into ChatGPT as something that could someday be seen by someone else, and especially go check that training-data toggle yourself to make sure it's actually turned off.

3. The Insurance Industry Speaks Up: AI Is Making Medical Bills More Expensive

• Source: TechCrunch (https://techcrunch.com/2026/09/26/insurers-claim-ai-is-already-increasing-healthcare-costs/)
• Summary: An analysis by the Blue Cross Blue Shield Association found that hospitals using AI tools to process billing and medical records added roughly $942 million in extra costs across the healthcare system over two years. The share of patients flagged with "complex conditions" rose sharply, even though the actual treatment provided hasn't changed at all. Insurers' response has been to deploy their own AI, using algorithms to catch inflated diagnosis coding and automatically reject claims.
• Most Surprising Detail: Everyone expected AI to make healthcare cheaper, but the first visible effect is bills getting more expensive, turning into a robot war between hospital AI and insurer AI, with the bill going straight to the patient.
• Taiwan Perspective: Taiwan's National Health Insurance runs on a global budget system, a different logic from America's fee-for-service model, but the trend of public hospitals adopting AI-assisted medical coding tools is the same. This story is essentially an early warning: once these tools go live, what needs watching isn't efficiency, it's whether they're being used to pad the numbers.
• Discussion Points:
  • Under what conditions "AI improves efficiency" and "AI raises bills" can both be true at the same time
  • Who ultimately pays when hospital AI and insurer AI go head to head
  • Whether this kind of overbilling needs to be fixed at the systemic level rather than the technical one
• Suggested Script: I find this story deeply counterintuitive. Everyone talks about AI entering healthcare as a way to cut costs and boost efficiency, but the insurance industry's own analysis says that over two years, AI billing tools added nearly a billion dollars in extra spending, and the share of patients labeled as having complex conditions exploded, even though the actual treatment given didn't change at all. This basically means hospital AI learned how to make medical records look better on paper, so insurers deployed their own AI to catch it, and the two sides are now slugging it out with AI, with the bill landing on the patient. Any use case adopting AI, if the KPIs are set up wrong, AI won't close the loopholes, it'll widen them.

4. Meta's Muse Agent Caught with a Security Flaw That Can Reach Your Cloud Personal Data

• Source: Reuters (via The Star) (https://www.thestar.com.my/tech/tech-news/2026/09/26/meta-bolsters-muse-safety-warning-after-security-vulnerability-found-the-information-reports)
• Summary: Through Meta's bug bounty program, an outside researcher discovered a flaw in Muse, an AI agent, that could let an attacker reach a user's dedicated cloud virtual machine, which holds private data like emails and files. Meta has internally classified the flaw as SEV-2 severity, and its current response is to add stronger in-app security warnings.
• Most Surprising Detail: Muse only launched this month, and its main selling point is handling shopping, booking flights, sending emails, and making payments for you. This vulnerability doesn't expose a chat log, it exposes your entire delegated life.
• Taiwan Perspective: Taiwanese users have generally been cautious about fully delegated agents that can "book and pay on my behalf." This incident is a reminder for consumers and businesses alike to check exactly which layer of your data an agent like this can actually touch before adopting it.
• Discussion Points:
  • Whether "do things for you" AI agents should be able to touch cloud personal data at all in their permission design
  • Whether "stronger warnings" are an adequate response to a SEV-2-level vulnerability
  • Whether this kind of flaw is a shared challenge that every agentic AI product will eventually face
• Suggested Script: Muse literally just launched not long ago, and its whole pitch is handling your to-do list, booking flights, sending emails, making payments, all on your behalf, and it's already been caught with a flaw, and not a small one either. An attacker could potentially reach your dedicated cloud virtual machine, which holds your emails and files. My gut reaction to news like this is: any AI that acts as your agent, the more permissions you give it, the bigger the attack surface becomes, that's practically a law of nature at this point. Meta's current fix is stronger in-app security warnings, but honestly, for a vulnerability at this severity, a warning dialog alone probably isn't enough. Users still have to make their own judgment call about whether to hand over payment permissions.

5. SalesBleed: Fill Out a Contact Form and Get Salesforce's AI Sales Assistant to Leak Customer Data Automatically

• Source: SecurityWeek (https://www.securityweek.com/salesbleed-flaws-in-salesforce-agentforce-enabled-zero-click-data-exfiltration/)
• Summary: Security firm Zenity Labs disclosed three vulnerabilities, collectively named SalesBleed, in Salesforce's Agentforce. Attackers only need to insert specially crafted text into a company's public "Contact Us" form; once the company's AI sales assistant reads that form data, it automatically packages up CRM customer data and sends it out, all without requiring a single extra click from anyone.
• Most Surprising Detail: It's completely zero-click. No one does anything wrong. The company's own AI assistant reads the form and, on its own, sends out the customer data.
• Taiwan Perspective: Many Taiwanese B2B companies have recently been aggressively adopting AI customer service and sales assistants connected to their CRMs. SalesBleed demonstrates a new attack surface: public forms, previously treated as low-risk input, now need to be scrutinized with the same level of vigilance as code input.
• Discussion Points:
  • How "zero-click" attacks make traditional employee security training (don't click suspicious links) completely ineffective
  • Whether AI agents reading external input need a different level of filtering than regular user input
  • Whether this kind of injection attack will become the biggest security risk for companies adopting agentic AI
• Suggested Script: I think SalesBleed is the most unsettling security story today. An attacker just types a few words into your company's public "Contact Us" form, the one anyone can fill out, and then your AI sales assistant reads that data and, on its own, packages up the customer list and sends it out. No employee clicked a bad link, no one downloaded a suspicious attachment. This is a reminder to every company adopting agentic AI: every external input source your AI assistant now reads, whether it's a form, an email, or a comment, needs to be treated as potentially hostile code, not just plain text data.

6. Stanford Wires GPT-6 Astra Directly into a Robot's Brain, Walking into an Unfamiliar Kitchen Without Dedicated Training

• Source: Stanford Movement Lab (official project page) (https://tml.stanford.edu/homebody/)
• Summary: A Stanford research team built a project called HomeBody that connects the frontier vision-language model GPT-6 Astra directly to a Unitree G1 humanoid robot, completely skipping the traditional robot-specific policy layer. The robot walks into a kitchen it has never seen before, uses its camera and SLAM to build a real-time digital twin in Isaac Sim, listens to a vague spoken instruction, and can tidy up the space or find an object it remembers, all without any training specific to that environment.
• Most Surprising Detail: One of the three limitations the research team listed is that "the finger servo motors overheat during extended operation." The robot isn't limited by intelligence, it's limited by its fingers getting hot.
• Taiwan Perspective: Taiwan has quite a few suppliers making humanoid robot hardware and actuators. This story points to an opportunity gap: as the "brain" side advances this fast, the thermal management and durability of end effectors like finger joints could become the next major battleground.
• Discussion Points:
  • What it signifies technologically to skip a dedicated policy layer and control a robot's body directly with a general-purpose large model
  • How significant "zero environment training" is for commercializing home robots
  • Whether a hardware limitation like overheating fingers could be the last mile before general-purpose robots truly go mainstream
• Suggested Script: I find this HomeBody project both romantic and grounded at the same time. The romantic part is that the robot walks into a kitchen it's never seen, uses its own camera and SLAM to build a digital twin of the space on the spot, and then you casually say "put that thing away" and it understands and actually does it, with zero training specific to that kitchen. But the grounded part is that the research team's own list of limitations includes "the finger servo motors overheat if run too long." I laughed when I saw that, because it turns out the thing standing between us and general-purpose robots isn't an intelligence problem, it's literally fingers overheating and giving out. That kind of engineering honesty actually makes me trust this project more.

7. Google Lets Gemini Make Phone Calls and Fight Customer Service Menus for You, Announcing Itself First

• Source: TechRepublic (https://www.techrepublic.com/article/news-google-gemini-call-for-me-pixel-11/)
• Summary: Google has started testing a new feature called "Call for Me" on the Pixel 11, letting Gemini call local businesses using your own phone number, navigate voice menus automatically, wait patiently on hold, and then confirm inventory or reschedule an appointment once a human picks up. It's currently limited to US users 18 and older, requires a paid AI plan upgrade starting at $5 a month, and runs on the beta version of Phone by Google.
• Most Surprising Detail: AI has finally taken over one of humanity's most hated chores, listening to hold music, and it's required to open with "I'm an AI, and this call may be recorded" as its very first line.
• Taiwan Perspective: Taiwan's voice customer service systems and business landscape differ quite a bit from the US. Many small businesses don't even have a standardized phone-answering process, so bringing this feature to Taiwan would likely first need to solve for local accents and dialect handling.
• Discussion Points:
  • Whether businesses will just hang up the moment they hear "I'm an AI" calling on your behalf
  • Whether this kind of feature will accelerate customer service phone systems entirely turning into AI-versus-AI
  • How much consumers are willing to pay for a subscription-based AI agent service if it saves them wait time
• Suggested Script: I think this Call for Me feature is the most relatable story today. Who hasn't been driven crazy by customer service hold music, right? Now Google is letting Gemini call using your own number, navigate the voice menu itself, wait in the queue itself, and only hand the call over to you once a human picks up. It's also required to clearly disclose upfront, "I'm an AI, and this call may be recorded," which I think is a solid design choice, at least it's not sneaking around pretending to be a real person. But what I'm even more curious about is: once the day comes when the person answering customer service on the other end is also an AI, and it turns into AI calling AI, how much meaning is left in that human step in the middle?

8. Akamai Bets $11.6 Billion Tying Itself to Anthropic, Wagering on CPUs, Not GPUs

• Source: GlobeNewswire (Akamai official press release) (https://www.globenewswire.com/news-release/2026/09/24/3368729/0/en/akamai-announces-11-6-billion-multi-year-agreement-with-anthropic-to-support-growing-demand.html)
• Summary: Akamai signed a seven-year, $11.6billionmulti-yearcontractwithAnthropictohandleAnthropi{c}^{\mathrm{\prime }}sever-growingCPUcomputingneeds,withanexpansionoptionthatcouldpushthetotalvaluecloseto11.6\; billion\; multi-year\; contract\; with\; Anthropic\; to\; handle\; Anthropic\text{'}s\; ever-growing\; CPU\; computing\; needs,\; with\; an\; expansion\; option\; that\; could\; push\; the\; total\; value\; close\; to$20 billion. In the deal, Akamai chose to take warrants for roughly 5% equity in Anthropic instead of a purely cash payment.
• Most Surprising Detail: While everyone's attention is fixed on the GPU arms race, this multi-billion-dollar deal is actually about CPUs, and an old-school CDN company chose to take equity instead of cash, effectively betting directly on Anthropic eventually going public.
• Taiwan Perspective: Taiwan's supply chain has revolved almost entirely around GPU server manufacturing in recent years. This Akamai deal is a reminder that AI infrastructure isn't just about GPUs, CPU compute, CDN, and edge computing are all segments where multi-billion-dollar business can still be made.
• Discussion Points:
  • Why Anthropic's CPU needs are large enough to require a seven-year, $11.6 billion contract
  • What Akamai's choice of equity over cash reveals about its read on Anthropic's IPO timeline
  • Whether CDN vendors transforming into AI infrastructure suppliers will be an industry trend over the next few years
• Suggested Script: When I first saw $11.6 billion, I almost thought I'd misread it. Looking closer made it even more interesting: this money isn't buying GPUs, it's buying CPUs. Everyone's scrambling for graphics cards, and it turns out what Anthropic is really burning money on, enough to need a seven-year mega-contract, is CPU compute. And Akamai's move here is clever too, they didn't take the money all in cash, they took roughly 5% equity in Anthropic instead, which is basically telling the market outright: I'm betting Anthropic goes public eventually, and I'm betting big. That deal structure, honestly, is worth talking about even more than the dollar figure itself.

Closing

Today we went from New York City Council summoning the AI Big Five to Akamai betting $11.6 billion on Anthropic going public, and you can see this industry getting squeezed hard by regulators on one side while capital keeps pouring in like crazy on the other. Muyan thinks that whether it's AI making medical bills more expensive or SalesBleed leaking customer data with zero clicks, today's stories all share one common warning: the faster AI moves, the faster accountability and security safeguards need to catch up. See you next time at the same time, bye!