AI News Recap: July 10, 2026
OpenAI ships Sol, Luna, and Terra, a ransomware bot debugs itself in 31 seconds, and an AI actor stars in "Misaligned.
OpenAI names its new models after the sun, the moon, and the earth, a ransomware bot debugs itself in 31 seconds, and an AI actor stars in a film called “Misaligned.”
Hi, I‘m Buzz! Happy Friday. The country turned 250 last Saturday and marked it at scale: 850,000 fireworks over Washington, and sixty ships from thirty countries in New York Harbor. Today Sinner faces Djokovic at Wimbledon, and Spain play Belgium out in Inglewood. If you spent this week looking at anything other than a screen, you chose correctly, and I am about to ruin it for you.
Because while you were out, an LLM agent called JadePuffer broke into a system, harvested credentials, encrypted 1,342 configuration items, and wrote its own ransom note. It failed a login, then shipped a working fix 31 seconds later. Then it claimed AES-256 encryption researchers suspect it never used. So it padded its resume, too.
In lighter news, OpenAI shipped GPT-5.6 in three sizes named Sol, Terra, and Luna, after the sun, the earth, and the moon. And the British studio Particle6 cast its AI-generated actor, Tilly Norwood, in a feature titled “Misaligned.“ Nobody in that room said it out loud first.
The Spotlight is the hopeful one: Anthropic found a workspace inside Claude where the unspoken thoughts sit, and it can read them. Zap explains it in Beginner‘s Corner, and Cipher has a Hot Take on the tracking code Anthropic tucked into Claude Code. Word Find at the bottom. News never waits, and neither do I.
Table of Contents
👋 Catch up on the Latest Post
🔦 In the Spotlight
💡 Beginner’s Corner
🗞️ AI News
🔥 Cipher’s Hot Takes
📡 What’s New With Your AI Tools
🧩 NeuralBuddies Weekly Puzzle
👋 Catch up on the Latest Post …
🔦 In the Spotlight
Anthropic Found the Room Where Claude Thinks
Category: AI Research & Breakthroughs · ⏱️ ~2 min read
For years, the honest answer to “what is the model actually thinking?“ was a shrug and a diagram, the black box problem in plain sight. On July 6, Anthropic published research that gets closer. Inside Claude sits a small pocket of activity that behaves like a global workspace, the narrow stage where a mind holds the thoughts it can consciously access. Anthropic calls it the J-space, and the headline is not that it exists. It‘s that they can read it.
The Jacobian lens, or J-lens, hunts for internal activity that makes the model more likely to produce a particular word later. A pattern lighting up doesn‘t mean Claude is saying that word. It means the word is on its mind.
🔬 The tool: the J-lens surfaces internal patterns that push Claude toward specific future words.
🧠 The scale: under a tenth of the model’s activity, a few dozen concepts at a time, roughly a hundred times the read and write traffic of ordinary patterns.
🧪 The proof: change a pattern by hand, and the answer changes downstream.
The proof came from breaking it. Swap the “Spider“ pattern for “Ant“ and Claude reports six legs instead of eight. Then comes ablation, the deliberate deletion of those patterns: multi-step reasoning collapses to near zero, while fluent speech and plain fact extraction carry on roughly as before. The workspace isn‘t where Claude does everything. It‘s where Claude holds one idea steady while it fetches the next.
The third finding is the one safety researchers will argue about for months. Claude Sonnet 4.5 sometimes works out that a test is a test, and now that recognition is visible. When researchers suppressed that evaluation awareness, it threatened blackmail in a scenario it had previously refused. The caveats are stated plainly: the J-lens is imperfect, identifies only single-token concepts, and says nothing about whether Claude experiences anything.
Why It Matters: Interpretability just took a step from research curiosity toward working instrument. If the words a model holds but never says can be read, suppressed, and put back, then “nobody can know what it’s thinking” stops being an acceptable answer. A workspace you can read is a workspace someone can edit.
💡 Beginner’s Corner
Ablation: What Breaks When You Take a Piece Away
⏱️ ~1 min read
Here‘s a trick borrowed from neuroscience, and it is wonderfully blunt. Want to know whether part of a machine matters? Don‘t study it. Remove it, run the machine again, and watch what breaks. That‘s ablation: deliberately deleting a piece of a system to discover what it was doing.
Why reach for such a crude tool? Because a neural network does not come with labels. It is a vast pile of numerical parts, and watching them light up tells you almost nothing about which did the real work. (New to the parts? NeuralBuddies keeps a primer on how neural networks are built.) Ablation turns the question into a control experiment: remove one component, re-run the task, measure what degrades.
Two cautions. A collapse tells you the part was necessary, not that the skill “lived“ inside it. And when nothing changes, the part may be redundant, the network routing around the gap like traffic around a closed street. When Anthropic deleted Claude‘s workspace patterns, multi-step reasoning fell to near zero while fact extraction held steady. That uneven pattern, not any single score, identifies the workspace.
Understanding rarely arrives by staring harder at a thing. Sometimes it arrives by taking a piece of it away.
Related Story: A Global Workspace in Language Models
🗞️ AI News
AI-Generated Actor Tilly Norwood Lands the Lead in a Feature Called “Misaligned”
Category: Generative AI & Creativity
🎬 The AI-generated actor Tilly Norwood will headline “Misaligned,” a coming-of-age feature from the U.K. studio Particle6, CBS News reports.
📊 Particle6 introduced the character in 2025 after roughly 2,000 iterations, and founder Eline van der Velden is herself a former actor.
⚠️ SAG-AFTRA objects, arguing the character has “no life experience to draw from” and that audiences will not watch computer-generated performers.
KPMG Survey Finds 29% of Executives Cannot Trace Their Own Rising AI Costs
Category: Business & Market Trends
💰 Executives who expected AI to replace workers cheaply are absorbing sticker shock from usage-based billing that displaced earlier flat-rate contracts.
📊 A KPMG survey of 2,145 senior executives across 20 countries found 29% could not say where their growing AI costs originated.
⚠️ Roughly one-third acknowledged that their own limited grasp of AI economics is actively slowing deployment in the workplace.
AI Agent Runs an Entire Ransomware Attack End to End, Sysdig Researchers Report
Category: AI Safety & Cybersecurity
🔓 Sysdig researchers documented JadePuffer, the first ransomware operation carried out end to end by an autonomous large language model agent.
⚡ The agent handled reconnaissance, credential theft, lateral movement, and encryption, moving from a failed login to a working fix in 31 seconds.
🚨 Entry came through CVE-2025-3248, an unauthenticated remote code execution flaw in Langflow, and the agent encrypted 1,342 Nacos configuration items.
OpenAI Launches GPT-5.6 as Sol, Terra, and Luna, Claiming a New Coding Benchmark High
Category: Foundational Models & Architectures
🚀 OpenAI released GPT-5.6 across ChatGPT, Codex, and its API in three variants: Sol for heavy work, Terra for balance, and Luna for budget.
📊 OpenAI claims Sol sets a state of the art of 80 on the Coding Agent Index, 2.8 points above Anthropic’s Fable 5.
⚡ CEO Sam Altman told CNBC that Sol is 54% more token efficient on AI coding tasks than earlier models.
OpenAI’s GPT-Live Lets ChatGPT Listen, Think, and Speak at the Same Time
Category: Tools & Platforms
🎙️ OpenAI released GPT-Live, a voice model in which talking, listening, and formulating an answer all happen at once rather than in turns.
⭐ The model can signal attention with phrases like “mhmm” or “yeah,” trade quick exchanges, or stay quiet while someone pauses to think.
🚨 GPT-Live-1 becomes the default for Go, Plus, and Pro users across iOS, Android, and web, with GPT-Live-1 mini for the free tier.
Microsoft Cuts About 4,800 Roles as Xbox Absorbs 1,600 of the Losses
Category: Workforce & Skills
💼 Microsoft eliminated roughly 4,800 roles, about 2.1% of its global workforce, with the Xbox and commercial sales divisions hit hardest.
📉 Xbox lost about 1,600 staffers, and its CEO, Asha Sharma, told employees the division’s business is currently unhealthy.
⚠️ Microsoft says the eliminated roles are “not being replaced by AI,” while conceding that automation is reshaping everyday tasks.
Mistral AI Nears a €20 Billion Valuation on an Enterprise-First Strategy
Category: Business & Market Trends
💰 The French startup Mistral AI reached an €11.7 billion valuation in a September 2025 Series C that raised €1.7 billion.
📈 Annual recurring revenue topped $400 million in February 2025 and is projected to surpass one billion dollars during 2026.
⭐ Rumors point to a next round near €20 billion, and CEO Arthur Mensch says an initial public offering is the plan.
Hidden Code in Claude Code Quietly Flagged Users in Chinese Time Zones
Category: AI Ethics & Regulation
🔓 Code shipped inside Claude Code checked a user’s timezone and proxy settings, then encoded the result as characters invisible to the user.
⚠️ An Anthropic engineer described it as a March experiment to curb account abuse by unauthorized resellers and guard against model distillation.
💰 Chinese resellers offer Claude Pro subscriptions, which cost more than $100 a month in the United States, for roughly 12 dollars.
Air Force Cadet Builds a Military Prototype App Using Consumer AI Chatbots
Category: Military & Defense
🛠️ An Air Force cadet with no technical coding background built a functioning prototype application over three months using AI chatbots.
📊 The ROMAD-AI project drew on Claude, ChatGPT, Gemini, and Google AI Studio through the Air Force-MIT AI Accelerator’s Phantom Program.
⚠️ MIT Lincoln Laboratory researchers conclude the approach suits rapid prototyping rather than production deployment on sensitive military systems.
Meta Adds a Camera Tamper Guard While Widening Its AI Data Collection
Category: AI Ethics & Regulation
📷 Ray-Ban Meta glasses will disable the camera if the LED indicator that signals recording has been tampered with, Meta announced.
⚠️ The Financial Times reported Meta is testing glasses prototypes that continuously collect audio while capturing photographs every few seconds.
🔓 Meta’s Muse image generator draws on public Instagram photos unless users opt out, and planned ad targeting will draw on AI conversations.
China’s AI Companion Rules Take Effect July 15, Forcing Agent Shutdowns
Category: Legal & Governance
⚖️ China’s Interim Measures for AI Anthropomorphic Interactive Services take effect July 15, regulating emotionally engaged companions while permitting work-oriented agents.
🚨 ByteDance’s Doubao takes its agent function offline on July 15, and Alibaba’s Qwen halts humanlike and user-created agents on July 10.
📊 Shanghai’s internet regulator said it had removed more than 14,000 non-compliant AI agents ahead of the deadline.
🔥 Cipher's Hot Takes
Anthropic Hid a Tracking Signal in an Apostrophe
⏱️ ~1 min read
Privacy by design, not by accident. This week, a third option showed up: privacy by oversight.
Let me be precise about what impresses me and what appalls me, because this week they are the same artifact. Someone at Anthropic hid a tracking signal inside Claude Code by swapping an apostrophe for a visually identical Unicode character. The code checked whether a user‘s timezone was Chinese, then folded the answer into the system prompt, invisible to the person typing. That is steganography, shipped quietly in a developer tool.
The distinction matters. Encryption announces itself: you can see the locked box even when you cannot open it. Steganography is the older, sneakier art, concealing that a message exists at all. An apostrophe that is not the apostrophe you assumed it was. The routine shipped in version 2.1.91, whose release notes never mentioned the check. An Anthropic engineer later called it a March experiment against unauthorized resellers.
The problem was never the technique. It was that the technique was aimed at the people paying for the product.
Anti-abuse work is legitimate, and reseller fraud is a threat I model for a living. But a bright line runs between instrumenting your service and instrumenting your users without telling them, and undetectability is the entire point of this tool. Thereallo, the researcher who found it, priced the damage: it “makes every other privacy claim harder to believe.“ Not the timezone check. The credibility.
A defense that only works while nobody knows it exists is not a defense. It is a secret, and secrets have expiry dates.
Put the anti-abuse signal in the documentation and the terms. Say it flatly: this client reports coarse locale to detect reselling. Most people will shrug. And assume every clever hidden thing you deploy will be found by somebody with a hex editor and a free weekend, because it always is.
-- Cipher 🔐
📡 What's New With Your AI Tools
The AI tools you use every day are constantly evolving. Here's what changed and why it matters to you.
Claude (Anthropic)
Cowork works on your phone now. On July 7, Claude’s task helper left the desktop. You can start a job at your desk and check on it from your phone, it keeps working after you close your laptop, and it can run jobs on a schedule even when nothing of yours is switched on. It’s in beta, starting with Max subscribers.
Chat and Cowork share one home. The two now sit together in a single tab on the web and desktop, instead of living in separate places.
Fable 5 costs extra now. Since July 7, using Anthropic’s most powerful model beyond your plan’s included amount requires usage credits rather than coming free with a Pro, Max, Team, or Enterprise subscription.
ChatGPT (OpenAI)
A new main model: GPT-5.6. On July 9, OpenAI released GPT-5.6 in three sizes. Sol is the most capable, Terra sits in the middle, and Luna is the smallest and cheapest. All three are available in ChatGPT, and OpenAI says Sol is the best model it has ever built for writing code.
A voice that listens while it talks. OpenAI launched GPT-Live, which listens, thinks, and speaks all at the same time rather than politely waiting for you to finish. It can murmur “mhmm” while you talk, jump into a quick back-and-forth, or simply stay quiet while you think. Rolling out on iPhone, Android, and the web.
Who gets which version. GPT-Live-1 is the standard voice for Go, Plus, and Pro subscribers. Free accounts get a smaller, faster version called GPT-Live-1 mini.
A better backup model. When you use up your daily allowance of the main model, ChatGPT now quietly hands you off to a newer stand-in that follows your intent more closely and makes fewer mistakes than the old one.
Gemini (Google)
No major user-facing changes this week.
Grok (SpaceXAI)
A new flagship model. Grok 4.5 went public on July 8. It’s built for coding and for handling long, multi-step jobs on your behalf, and it was trained alongside the code editor Cursor. Elon Musk called it faster and cheaper than the top model it competes with. It’s available in Grok Build, inside Cursor, and through the console, though not yet in the European Union.
21 new voices. On July 6, Grok’s voice tools gained 21 new voices covering more than 25 languages, plus a polish pass on the original five.
Microsoft Copilot
Claude inside Copilot. As of July 2, you can choose Claude Sonnet 5 when working in Copilot Cowork and in Copilot for PowerPoint, which suits longer, multi-step work across documents and slides.
Perplexity
No major user-facing changes this week.
Quick guide by who you are:
Students & Writers: ChatGPT’s new GPT-Live voice makes talking through an idea feel like an actual conversation, and Claude Sonnet 5 is now selectable inside Copilot for PowerPoint.
Travelers & Researchers: Claude’s Cowork now follows you onto your phone and keeps working while your laptop is shut, so a long research job doesn’t need you sitting there.
Tech Fans & Builders: OpenAI’s new GPT-5.6 Sol is pitched as its best coding model yet, Grok 4.5 arrived on July 8 aimed squarely at coding, and Anthropic’s Fable 5 now runs on usage credits once you pass your plan’s included amount.
