In partnership with
In Today’s Issue:
🧪 Hackers reach patient data at Quest Diagnostics
💰 OpenAI nears a $1.25 trillion round led by UAE money
🔓 MiniMax opens up its M3.1 Flash weights
📱 Grok Bot learns to use your phone's apps
🐳 Washington orders DeepSeek off federal computers
✨ And more AI goodness…
⚡ The Signal
This week the AI security race stopped being abstract.
Earlier in the week, AI-discovered flaws in PHP, the language behind about 70% of websites (we reported yesterday), sent companies into an emergency patching scramble. Today Quest Diagnostics said hackers had reached patient data in a portal clinics use to look up lab results. Quest says the two are unrelated, but together they show the pressure CrowdStrike's George Kurtz describes in today's video: attackers armed with AI agents now move in minutes, sometimes seconds. The same speed is arriving on defense. Anthropic just showed Claude proving security properties across more than 100 open-source libraries in a few weeks, and Washington is ordering DeepSeek off federal computers over where government data could end up. The edge goes to whoever puts AI to work protecting systems first.
All the best,
Kim Isenberg
(Getty Images via CNBC)
💰 OpenAI Nears a $1.25 Trillion Round Led by UAE Money
OpenAI is close to raising up to $30 billion at a $1.25 trillion valuation before the new money, with investors from the United Arab Emirates leading alongside Coatue, CNBC reported Thursday, citing people familiar with the deal. OpenAI wants signatures by the end of next week and is also negotiating a deal for more computing capacity from the UAE. The round is being positioned as an "opportunistic" bridge before a possible IPO; OpenAI filed its prospectus confidentially in June.
👉 tl;dr: OpenAI is lining up one more giant private round, plus Gulf computing power, before it lists on the stock market.
(Hugging Face)
🔓 MiniMax Keeps Its Promise and Opens M3.1 Flash
MiniMax published the weights of M3.1 Flash on Hugging Face early Friday, as it had promised when the agent model launched on Monday. Anyone can now download the model and run it on their own hardware: 428 billion parameters in total, 23 billion active per token, with up to 1 million tokens of context. MiniMax pitches it for coding, agents, computer use and cybersecurity, citing its own scores such as 82.2 on CyberGym. Non-commercial use is free; commercial users must credit MiniMax and notify it, and companies above $20 million in yearly revenue need written approval.
👉 tl;dr: The low-cost Chinese agent model we charted on Tuesday is now downloadable, so companies can run it on their own servers instead of MiniMax's cloud.
(Grok Bot on X)
📱 Grok Bot Can Now Work Inside Your Phone's Apps
Grok Bot, Elon Musk's AI assistant, can now use Calendar, Reminders, Contacts and other apps on iPhone and Android, rolling out now. Its permission screen lists Messages (read and send), Location, Calendar, Home controls such as lights and locks, Reminders and Contacts, each with its own Allow button. A 1Password integration announced the same day, arriving over the next few days, lets the bot sign in to sites with logins you approve. It is the kind of cross-app help Apple promised for Siri back in 2024 and then delayed. Crazy when you think about it.
👉 tl;dr: Grok Bot can now text, schedule and organize for you inside your phone's own apps, once you grant each permission.
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. No rough transcription you have to clean up. Actual polished text, ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
🎬 Watch This
CrowdStrike CEO George Kurtz tells No Priors that "the new apex predator is the agent state, not the nation state." Open-weight models with their guardrails removed now hand a low-level activist the toolkit a government hacker had a few years ago, he argues, and attacks run on prompts instead of malware. Once inside a network, intruders now move to the next machine in 29 minutes on average and in as little as 27 seconds. His answer fits today's lead: defenders need frontier-grade AI of their own. Jump to 2:54 for his "pyramid of pain" and 7:34 for breakout times.
"I’m confident we can collectively navigate the potential challenges of frontier models and usher in a new golden age of scientific discovery!"
– Demis Hassabis, CEO of Google DeepMind, on X
Hassabis posted it Thursday after the WSJ featured him in its Innovators issue. In a week of AI-found software flaws and a patient-data breach, the Nobel laureate who runs Google DeepMind is betting that the upside wins.
Anthropic appears to be quietly testing new Claude Max limits. SemiAnalysis says some Max 20x accounts now get about 20% more usage per five-hour window but about 20% less value per month, which would explain users' complaints about lower limits.
🧪 Hackers Reach Patient Data at Quest Diagnostics
The Takeaway
👉 Quest Diagnostics says hackers accessed Quanum Lab Services Manager, a portal some clients use to order tests and look up results; it found the intrusion Wednesday and cut off access.
👉 Patient health and identity data were involved; how many people and which records is still under investigation.
👉 Quest told the AP the breach is not related to the PHP2Shell flaws that shook the web this week, and no attacker has been named.
👉 Defense is speeding up too: Anthropic used Claude to prove security properties in more than 100 open-source libraries in a few weeks.
Hackers have reached patient data at Quest Diagnostics, the lab company that says it serves half the physicians and hospitals in the United States. On Thursday, Quest disclosed "unauthorized access" to Quanum Lab Services Manager, a portal certain clients use to place test orders and look up results. It found the intrusion on Wednesday, shut the system off and brought in law enforcement and outside security experts. Asked whether patient health or identity data was involved, Quest's notice answers: "Yes, but we are still investigating to determine the specific data affected."
(Quest Diagnostics notice, via Andrew Curran on X)
It is close to the worst case for any breach: you can change a password afterwards, but not your lab results. It also lands in the same week as the PHP2Shell scramble we covered yesterday, when AI-discovered flaws in software behind about 70% of websites drew millions of exploit attempts. Quest told the Associated Press its breach is unrelated, and no attacker or method has been named. Quest patients have been here before: a 2019 hack at a billing vendor exposed data on about 11.9 million of them.
(Quest Diagnostics notice, via Andrew Curran on X)
The defense is getting faster too. On Thursday, Anthropic showed Claude proving security properties in more than one hundred real-world open-source libraries over a few weeks, work that until recently was feasible only for small codebases or with a lot of expert time. As CrowdStrike's George Kurtz puts it in today's video: "The only way to fight AI today is with better AI."
Why it matters: Medical records stay sensitive for life, and one shared portal can expose patients across many clinics at once. With AI making attacks faster and cheaper, the companies that hold health data now need defenses that work at the same speed.
Sources:
🔗 https://t.co/B3mouySJeE
🔗 https://x.com/AndrewCurran_/status/2113437710065311970
🔗 https://www.cnbc.com/2026/10/22/quest-diagnostics-reports-cyber-breach-of-certain-patient-records.html
When your support agent gets it wrong, who's accountable?
Every agent handling refunds, account changes, or tickets needs three answers: what it can pick up, what it's allowed to do, and when it hands back. Join the live conversation Oct. 27, 10 a.m. PT, part of Running Agents in Customer Work. Register today, recordings included.
(The HIPAA Journal, based on HHS Office for Civil Rights data)
The chart: How many people had health records exposed in hacking and IT incidents each year, counted by The HIPAA Journal from breach reports filed with the US health department (breaches affecting 500 or more people). The area is near zero until 2014, spikes to about 110 million in 2015, the year of the Anthem hack, then peaks at roughly 270 million in 2024, when the Change Healthcare ransomware attack alone hit 192.7 million people. 2025 lands around 135 million; 2026 already passes 70 million with data through August 25.
The lesson: The biggest losses now come from the hubs in the middle of American medicine: clearinghouses, billing firms and shared portals that connect thousands of providers. Change Healthcare and Conduent (62.2 million people in 2025) are such hubs, and a lab-ordering portal like Quest's plays a similar connecting role.
The caveat: One mega-breach can swing a whole year, smaller incidents are not published, and 2026 is incomplete. Quest has not said how many patients were affected, so today's breach is not in this chart yet.
🐳 Washington Orders DeepSeek Off Federal Computers
⚡ Bottom line
DHS has ordered civilian federal agencies to stop buying DeepSeek and remove it, including tools other vendors built on it.
💡 Why it matters
The order reaches past DeepSeek's own app into products built on its open models, so contractors must know what powers their tools.
🔎 What it means
The US-China AI contest is moving from chips into software supply chains, and Beijing has already vowed to defend its companies.
Washington has told its civilian agencies to rip DeepSeek out of their computers, and not only DeepSeek's own app. The Department of Homeland Security order, published in the Federal Register on Tuesday, also covers "third-party products and services developed using DeepSeek implementations": any vendor tool built on the Chinese lab's models. Agencies must stop buying them at once and remove what they run, or ask for an exception within 60 days.
(DHS order via the Federal Register)
The order, signed by Homeland Security Secretary Markwayne Mullin, carries an original release date of July 24; it became public only this week. Its reasoning is about data: DeepSeek's servers sit in China, Chinese law gives state cyber and intelligence agencies broad access to company infrastructure, and whatever a government employee types into the chatbot could flow there. That, DHS writes, "could produce actionable intelligence on U.S. persons and/or government programs."
The open-weight clause reaches furthest. DeepSeek gives its models away, so many products run them without ever touching DeepSeek's servers. DHS reaches those too, warning that "the models themselves may contain malware or other security vulnerabilities."
(Getty Images via TechCrunch)
Beijing pushed back on Tuesday. Foreign Ministry spokesperson Mao Ning said China opposes "overstretching the concept of national security and politicizing technology and trade issues." The AI rivalry now runs through software supply chains as well as chips. As more Chinese labs publish their weights, as MiniMax did this morning, knowing which model sits inside a product becomes part of every government purchase.