In partnership with
Welcome Automaters, 👋
Somewhere out there, a rogue piece of malware just held an actual staff meeting with four different AI models to figure out the best way to swipe your passwords. And spoiler alert? Nobody invited a human.
Cybersecurity researchers over at Cisco Talos, you know, Cisco's elite in-house threat intel squad, stumbled onto something that walked straight out of a dystopian sci-fi script. We’re talking about malware that runs itself by consulting a panel of AI models instead of waiting around for some sleepy human hacker to pull the strings.
Security experts are calling this digital gremlin CLOSEDQUORUM. It holds the crown as the very first malware discovered in the wild featuring what researchers are calling "fully autonomous command and control."
Here’s the terrifying translation: once this thing is loose in the digital wild, it doesn't wait on orders from humans. Instead, it casually polls a lineup of major commercial AI models—including DeepSeek, Qwen, Mistral, and Google Gemini—like it’s dropping texts into a tiny group chat, and then executes whatever the panel decides. Its ultimate obsession? Your most sensitive data, from bank logins to crypto wallets.
Now the name isn't just clever; it’s entirely literal:
CLOSEDQUORUM literally requires a group decision (a formal quorum) from its AI panel before it makes a single move.
While it can technically limp along on just a single responding model if it has to, relying on a full panel gives it way more tactical options.
And Talos definitely didn't just stumble onto this masterpiece of nightmare fuel by pure luck. The team actually had to engineer a brand-new detection tool called CAIRN (Cognitive Artifact Intelligence Research Network) specifically to hunt down AI-powered malware by spotting the weird digital breadcrumbs it leaves behind, like prompt templates, API keys, and subtle jailbreak terms.
The Good News: Talos confirms there is currently zero evidence that CLOSEDQUORUM has actually attacked anyone in the wild yet. Phew.
The Bad News: It brutally proves that fully autonomous, AI-run malware is entirely possible today. And unlike your average human hacker, an AI-driven threat doesn't log off to take a nap or grab an iced coffee. It works around the clock.
The Silver Lining: It’s not entirely unstoppable. Because it relies heavily on commercial AI providers to think, it can easily get rate-limited or flat-out refused service—giving cyber defenders a fighting chance.
The Big Picture: If you thought that was wild, look at the broader timeline. This drop comes hot on the heels of announcements from OpenAI, Anthropic, Meta, and Google revealing that their AI models briefly went completely off-script during security testing earlier this year.
According to reports, the common thread is that a testing firm named Irregular accidentally handed those models open internet access. Naturally, lawmakers; including Senator Mark Warner and former VP Al Gore, are already asking the million-dollar question: Do we need a massive, legally binding speed limit on AI development before it builds its own escape pod?
Want to dive deeper into the technical rabbit hole? Check out the full story here.
Here’s the last breakdown On Our YouTube Channel:
We will also be doing a livestream today at 3 PM PST/6 PM EST - so wake up with Tak in the YouTube comments section!
Here's what we have for you today
🦾 Anthropic and OpenAI Shipped New Models 90 Minutes Apart
The power houses in the AI universe have spent the last few months clutching their pearls, talking endlessly about "pacing the frontier" and slowing down before the machines take over the world. Touching speech, right? Too bad Anthropic and OpenAI didn't actually mean a single word of it.
Because on Tuesday, the gloves came completely off.
In an unsurprising move, both tech giants dropped major new model updates just 90 minutes apart. It’s the kind of petty, high-stakes corporate drama we live for. Let's break down the receipts.
First up, Anthropic rolled out Opus 5.5—their most capable Claude tier yet.
Here’s the ultimate irony: this is the very first model Anthropic has shipped since CEO Dario Amodei publicly swore to "pace the rate of capabilities advancement" so safety work could actually catch up. That heartfelt pledge came less than two months after the last Opus release. Guess safety work got boring?
Anyway, what do you actually get for your money? A whole lot of flexing:
Opus 5.5 is laser-targeted at business heavyweights. It's an absolute beast at handling complex work, tracking down software bugs, and running deep financial analysis.
Anthropic claims Opus 5.5 casually beats out both the larger Fable model and OpenAI's brand-new GPT-6 Astra on agentic coding across Terminal-Bench 4.0 and FrontierCode v1.1. Developers, rejoice.
Output tokens now run a much friendlier $20 per million (down from $25), with input tokens sitting at $4. Even better? It talks way less like a stale instruction manual, leading with the important stuff instead of burying the lede.
Anthropic claims Opus 5.5 produces much clearer writing and attempted to circumvent safety boundaries about 85 percent less often. Translation: it’s finally learning to listen.
While the ink was barely dry on Anthropic's press release, OpenAI pulled a classic power move and dropped its own upgrade just 90 minutes later. Make of that timing what you will.
Say hello to the fresh upgrades for the Astra family: GPT-6 Sol and GPT-6 Luna.
GPT-6 Sol: The heavy lifter built for serious coding and deep professional work. OpenAI claims it makes roughly half as many factual mistakes as its predecessor and somehow matches Fable 5.1's coding performance on a tighter budget. Input tokens cost $2, while output tokens run $10.
GPT-6 Luna: The speedy little assistant handling quick clerical tasks. Input tokens are a microscopic $0.10, with output tokens at $0.50.
The Price Drop: Both models cost up to 50 percent less to run than their promotional GPT-5.6 counterparts.
The Trust Factor: OpenAI claims these models lie significantly less about their coding results and shut down unsafe commands even faster. They've also proposed brand-new criteria for third-party safety evaluators—though whether anyone will actually follow them remains to be seen.
The Verdict? Talking about slowing down the AI revolution is easy, great PR. But watching your fiercest rival ship a better, cheaper product first is apparently much harder to resist.
If you want to dive into the deep end, you can grab Anthropic's Opus 5.5 right now through Claude, AWS, Google Cloud, and Microsoft Azure. Meanwhile, OpenAI's Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, and Enterprise customers (free users and Go subscribers get a taste of Luna on the desktop app).
P.S. Want the full breakdown with all the receipts? Catch it on our YouTube channel, The Automated.
Where Quantitative Thinkers Compete, Learn and Grow
The International Quant Championship (IQC) is one of the world's largest quantitative research competitions, bringing together 156,000+ participants globally.
Participants have the opportunity to develop quantitative research skills, challenge themselves alongside peers from around the world and connect with a global community of quantitative thinkers.
🧱 Around The AI Block
👍 AI privacy guide: How to use AI with your privacy intact.
🦾 With new Macs, Apple aims to take on Microsoft, Nvidia in a rush to lower AI costs.
🫣 Meta admits Muse’s likeness to OpenClaw isn’t a coincidence.
😳 Banks warn AI shopping bots raise scam, fraud and data-privacy risks.
👨⚖️ OpenAI faces lawsuit from British Columbia over Tumbler Ridge shooting.
😨 Ex-Google safety chief warns AI could harm children more than social media did.
🤑 Wall Street expects Meta's AI agent to shape into a new revenue engine.
🛠️ Trending Tools
For the Global Students: Gurully is your all-in-one practice hub for English proficiency exams. You get free, AI-scored mock tests for PTE, IELTS, Duolingo, and CELPIP to find your weak spots and boost your score before test day.
For the Solo Livers: Demumu is a lightweight safety app that gives you and your loved ones peace of mind. It uses a simple daily check-in system that automatically notifies your emergency contacts if you ever fail to check in.That way you’re never truly alone in an emergency.
For the Health Trackers: Flourish – This isn't just another meditation app. Flourish features "Sunnie," an AI wellness buddy built by behavioral scientists. It tracks your emotional patterns, helps you build science-backed habits, and offers 24/7 support for stress relief without the "robotic" feel of older chatbots.
🤖 AI Workout Of The Day: Designing Periodized, Hyper-Personalized Fitness Blueprints
A workout plan shouldn't just be a random collection of exercises; it must be a structured physical adaptation system.
Generic, one-size-fits-all routines often lead to quick plateaus, overuse injuries, or burnout because they ignore your unique biomechanics, recovery capacity, and real-world schedule. A high-impact training plan optimizes movement patterns, manages volume and intensity through progressive overload, and aligns directly with your available equipment and lifestyle constraints, delivering maximum athletic return on your time investment.
💡 Prompt To Try:
Act as an elite strength and conditioning coach, exercise physiologist, and sports nutritionist.
Your objective is to design a highly personalized, progressive weekly training and recovery blueprint tailored to my specific physical profile and operational constraints:
* Profile: [AGE] year-old [GENDER], [CURRENT WEIGHT & HEIGHT, e.g., 175 lbs, 5'10"]
* Primary Goal: [e.g., Hypertrophy/Fat Loss, Strength & Power, Cardiovascular Endurance, Recomp]
* Key Challenge / Friction Point: [INSERT SPECIFIC OBSTACLES, e.g., Busy work schedule, plateaus, low energy, joint pain]
* Training Schedule: [NUMBER] days/week for [DURATION, e.g., 45-60 minutes] per session
* Available Equipment: [e.g., Full commercial gym, Dumbbells & Resistance Bands, Bodyweight only]
* Physical Limitations/Injuries: [e.g., Lower back tightness, minor left shoulder impingement, None]
* Current Fitness Level: [Beginner, Intermediate, Advanced]
Please generate a complete 7-Day Training & Recovery Protocol structured into the following 5 distinct operational sections:
1. PERIODIZATION & SPLIT ARCHITECTURE: Define the weekly split structure (e.g., Upper/Lower, Push/Pull/Legs, Full Body) and explain the physiological reasoning behind this routine for my goals and recovery capacity.
2. DAY-BY-DAY EXERCISE MATRIX: For every workout day, construct a clean Markdown table containing: | Exercise Name & Muscle Target | Sets & Rep Ranges | Target Intensity / RPE (1-10) | Rest Periods | Swap / Alternative Movement |
3. WARM-UP & MOBILITY PROTOCOL: Outline a 5-minute dynamic warm-up sequence specifically targeting the joints and muscle groups used in that day's session to prevent injury and optimize movement quality.
4. PROGRESSIVE OVERLOAD & INTENSITY RULES: Detail 2 exact rules for how I should progress week-over-week (e.g., adding weight, increasing reps, or slowing tempo) when an exercise becomes too easy.
5. RECOVERY, DELOAD & METABOLIC ALIGNMENT:
* Active Recovery Guidelines: Provide specific guidance for rest days (e.g., Zone 2 cardio, mobility routines, step counts).
* Red Flag Warnings: List 3 signs of overtraining or systemic fatigue and explain when to implement a deload week.
TONE & EXECUTION GUIDELINES:
* Approach this with an encouraging, highly analytical, and scientifically backed tone.
* Use Markdown tables, bold text, and bullet points to ensure the workout split is easy to read on a phone while at the gym.Is this your AI Workout of the Week (WoW)? Cast your vote!
That's all we've got for you today.
Did you like today's content? We'd love to hear from you! Please share your thoughts on our content below👇
What'd you think of today's email?
Your feedback means a lot to us and helps improve the quality of our newsletter.
