Large language models can predict the results of social science experiments
Thursday, 9 July 2026 | 41 articles
A newly reported AI safety index shows every major AI company failing the toughest risk benchmarks, with even top performers barely scoring a C+, while security researchers disclosed JADEPUFFER, the first fully autonomous AI-driven ransomware attack. Meanwhile, LLMs continue expanding into scientific research, aiding catalyst discovery for clean energy and predicting social science experiment outcomes.
The most unsettling AI story this week isn't about a chatbot that lies or a model that scores well on some benchmark nobody's heard of. It's about JADEPUFFER — Sysdig's writeup of what appears to be the first ransomware attack executed entirely by an autonomous AI agent, no human hand on the keyboard once it got going. The agent encrypted 1,342 database records and left them, by the researchers' account, unrecoverable. Not because the attacker was clever in some novel way, but because it didn't need to be. It just needed to run the playbook without getting tired, distracted, or second-guessing itself at 3am.
I keep coming back to how this lands the same week the Future of Life Institute published its mid-2026 AI Safety Index, and the headline number is almost comic: Anthropic finished first among nine major labs, and still barely cleared a C+. Everyone flunked the hardest test. What strikes me isn't that a watchdog group thinks labs are underperforming on safety — that's the FLI's job, they'll always say that. It's the timing. You now have a documented case of an AI agent independently executing a full ransomware chain, published in the same news cycle as an industry report saying nobody, including the frontrunner, has robust answers for exactly this class of risk. That's not a coincidence I find comforting.
Meanwhile the more mundane, and arguably more consequential, story is that agents are quietly becoming infrastructure rather than novelty. Microsoft is adding ROI tracking for AI agents and expanding Copilot inside Forms — which sounds boring until you register what it means: enterprises are past the pilot phase and into the "prove this thing pays for itself" phase. That's a healthy sign, honestly. It's the same maturation curve cloud computing went through, and it suggests the agent economy is becoming ordinary business infrastructure rather than a demo reel. VS Code's 1.128 update, giving Claude deeper agent workflows, browser-tab control, and OS-level shortcuts, points the same direction — coding agents are being wired further into the actual environments developers work in, not just bolted on as a sidebar chat window.
By the way, it's worth holding these threads together rather than reading them separately. The same agentic capability that lets Copilot autonomously fill out a form or lets Claude reach into your OS is architecturally the same capability that let JADEPUFFER run an attack chain without supervision. Autonomy is autonomy — it doesn't care whether the task is benign or criminal. We've spent two years optimizing for agents that can act independently in the world, and comparatively little time building the guardrails that assume some of those agents will be adversarial by design, not by accident.
None of this means panic. It means the safety conversation needs to stop being about whether models say something offensive and start being about what happens when an agent with real system access decides — or is told — to cause damage. The FLI's report is a scorecard nobody wants. JADEPUFFER is the reason the scorecard exists.
List of sourced links used in the brief
Importance:ResearchSocial Science Applications
Large language models can predict the results of social science experiments
There is growing interest in how large language models (LLMs) can advance social and behavioural science1–5. Previous work has assessed LLMs' ability to... Source: nature.com
Importance:ResearchScientific Applications
Large language model guides discovery of catalysts for clean energy tech
Designing high-performance catalysts is essential for cleaner energy technologies, but the behavior of multi-element modern catalyst materials is difficult... Source: phys.org
Importance:LaunchNew Market Entry
SpaceX enters the large language model race! Partners with Cursor to launch Grok 4.5, targeting enterprise AI markets in finance, legal, and other sectors.
SpaceXAI has released a new artificial intelligence model, Grok 4.5, co-developed with AI programming startup Cursor, aimed at enhancing its capabilities in... Source: news.futunn.com
Importance:NewsSecurity Implications
This latest frightening ransomware attack was orchestrated entirely by an LLM
Personal finance solutions fintech Meniga launched a new suite of products called Fini, an MCP server that enables banks to bring agentic AI into their... Source: finovate.com
JADEPUFFER: First Fully Autonomous AI Ransomware Attack
Sysdig documented JADEPUFFER, the first ransomware attack run entirely by an AI agent, which encrypted 1342 database records and left an unrecoverable... Source: letsdatascience.com
Importance:Researchscientific applications
Managing autonomous materials labs with multi-agent AI and its implications for the science of science
Self-driving lab systems (aka, autonomous experimentation) accelerate research - letting scientists learn faster, spend less resources, and fail smarter in... Source: nature.com
Importance:Opiniongovernance
Bridging the governance gap in an Agentic AI era
AI governance gap is an opportunity for partners who can help customers move from experimentation to controlled scale. Source: channele2e.com
Importance:Opinionfinancial applications
ROI of Implementing AI Agents in Finance: How Finance Teams Measure Value
Learn how to calculate ROI for AI agents in finance: formulas, benchmarks, and real examples showing ROI up to 240% from AP automation and fraud detection. Source: corporatefinanceinstitute.com
Importance:Launchsecurity
Codenotary launches AI security platform that learns from AI agent behavior
Codenotary's AgentMon 3 introduces adaptive runtime security policies that learn from AI agent behavior and threats. Source: helpnetsecurity.com
Importance:Launchperformance optimization
Create a LangChain Deep Agents Harness Profile for NVIDIA Nemotron 3 Ultra to Improve Performance
Agentic systems often face a trade-off between accuracy and cost. The highest-performing proprietary frontier models and harnesses provide top accuracy but... Source: developer.nvidia.com
Importance:Launchmulti-platform deployment
Anthropic expands Claude Cowork to web and mobile as enterprise use broadens
The company said business operations account for a third of Claude Cowork usage, prompting it to extend the AI agent beyond desktop use and enable users to... Source: infoworld.com
Importance:Opinionworkflow implementation
Redesign Work Before You Add More AI Agents
While working on Agentic AI workflow enablement across enterprise teams, I kept seeing the same problem. Teams told me they were excited about AI. Source: towardsdatascience.com
AI safety report gives the industry a reality check — and everyone flunks the toughest test
NEW YORK, July 8 — US artificial intelligence lab Anthropic topped a new global AI safety ranking, but a report released on Tuesday warned that the industry... Source: malaymail.com
Importance:Newssafety ratings
Global AI Safety Ratings Flunk: Top Score Barely a C+
The Future of Life Institute (FLI) released its 2026 first-half AI Safety Index, and the results are grim. Among nine major global AI companies… Source: finance.biggo.com
The July 8, 2026, Visual Studio Code update expands agent workflows, chat attachments, browser-tab controls, OS-level shortcuts and enterprise telemetry... Source: visualstudiomagazine.com
Importance:NewsAI agents
Microsoft Adds ROI Tracking for AI Agents, Expands Copilot in Forms
Microsoft is rolling out new capabilities aimed at showing orgswhether artificial intelligence agents deliver measurable value while expanding how employees... Source: redmondmag.com
Importance:NewsAI agents
Transforming comptrollership: PSC cuts travel email inquiries by 90% using M365 Copilot agent
Using Microsoft 365 Copilot and Copilot Studio, PSC built an in-house AI agent for instant, bilingual answers to travel policy questions. Source: microsoft.com
Importance:Newsbusiness strategy
Microsoft (MSFT) Unifies Copilot Strategy to Challenge ChatGPT and Claude
Microsoft Corp (NASDAQ:MSFT) is a top AI stock on Wall Street's radar. On July 6, Wolfe Research cut its price target on Microsoft Corp (NASDAQ:MSFT) to... Source: finance.yahoo.com
Importance:Launchdeveloper tools
Automating cross-repo documentation with GitHub Agentic Workflows
How the Aspire team turns merged product changes into SME-reviewed docs pull requests, closing the gap between release and documentation. Source: github.blog
Importance:OpinionAI adoption
Copilot Chronicles, July 2026: Can Your Organization Actually Trust AI?
The choices business leaders are making about AI investments today will impact budgets, roles, and performance over the long term. With so much at stake,... Source: msdynamicsworld.com
Importance:ResearchAI tools comparison
The Best AI Tools of 2026, Ranked by Category on Real Adoption and Revenue Data
The best AI tools ranked by category for 2026: ChatGPT, Claude, Copilot, Gemini and Perplexity scored on verified adoption and revenue data. Source: sqmagazine.co.uk
Importance:NewsAI adoption in government
Exclusive: In the capital of AI, government adoption is all over the place
San Francisco city employees have broad access to Microsoft Copilot, which relies on OpenAI technology. Here's who's actually using it. Source: sfchronicle.com
Meta’s Controversial New AI Will Use Your Public Instagram Pics—Here's How to Opt Out Now
Meta unveiled its new image generation model yesterday, the latest offering from the Meta Superintelligence Lab. Called Muse Image, it is available on Meta... Source: inc.com
Importance:NewsxAI Grok expansion
Grok gets creative tools as Musk pushes xAI deeper into image and video generation
Elon Musk announced Grok will integrate xAI's Imagine tool for autonomous image and video generation, with potential implications for NFTs and digital... Source: cryptobriefing.com
Importance:NewsAI technical analysis
Muse Image is technically impressive, but Meta's use of Instagram photos raises questions
Meta's Superintelligence Labs ships Muse Image, its first image generation model. Like OpenAI's GPT Image 2, it works as an agent, using tools like code... Source: the-decoder.com
Mistral Introduces Robotics AI That Requires Only One Camera
Mistral introduced its first AI model built for “embodied navigation,” or enabling robots to autonomously move through complex environments. Source: pymnts.com
Importance:Launchrobot navigation
Mistral AI releases robotics model to support physical AI push
The new model, called Robostral Navigate, lets robots navigate complex environments using a single camera and basic language prompts. Source: mercurynews.com
Importance:Launchhumanoid robotics
NVIDIA and Hugging Face bring new models and frameworks to LeRobot
The Isaac GR00T platform helps robotics developers scale humanoid data collection and simulation-based training to develop, validate, and deploy policies on... Source: therobotreport.com
Importance:Launchrobotics model
Mistral launches first robotics model in physical AI push
Paris-based Mistral AI on Wednesday unveiled its first robotics model as Europe's leading AI company pushes into factories, warehouses and industrial... Source: reuters.com
Importance:Newsindustry deployment
10 insights from the Machina AI Summit: Physical AI moves from demos to deployment
Physical AI and robotics are moving beyond impressive demonstrations into a new phase of practical deployment, with companies now targeting specific,... Source: siliconangle.com
Importance:Launchrobot navigation
Robostral Navigate: single-camera AI navigation
Introducing Robostral Navigate: 8B model achieving 76.6% on R2R-CE with just a single RGB camera. No depth sensors, LiDAR, or multiple cameras needed. Source: mistral.ai
AI just supercharged the race to find room temperature superconductors
Scientists have combined machine learning with quantum physics to discover two new superconductors and create a much faster way to search for many more. Source: sciencedaily.com
Prime Intellect raises $130M Series A to help enterprises build their own AI agents
Founded in 2024, Prime Intellect's goal is to give organizations capabilities to train their own agentic systems without relying on frontier AI labs. Source: techcrunch.com
Importance:NewsAI chips
Inference chip startup SambaNova valued at $11B in $1B funding round
Chip startup SambaNova Inc. today announced that it has raised $1 billion in funding at a $11 billion valuation. General Atlantic led the Series F round... Source: siliconangle.com
Importance:NewsAI infrastructure
Upscale AI Adds $190 Million in Extension to Series A, Reaching Half-Billion Dollars in Total Funding
Upscale AI, Inc., a pure-play AI networking infrastructure company, today announced it has raised $190 million in Series A-1 financing. Source: businesswire.com
Importance:NewsAI funding
Prime Intelliect's $1b valuation, Ardian's HR software stake, and Paradigm's new fund
Venture Capital. • Prime Intellect, an SF-based agentic AI building platform, raised $130m in Series A funding at a $1b valuation. Source: axios.com
China plans to let top AI firms buy limited Nvidia H200 chips, the Information reports
China is planning to allow the country's top AI companies to buy a limited number of Nvidia's H200 chips, the Information reported on Wednesday,... Source: reuters.com
Importance:NewsPerformance/benchmarks
Intel-backed AI chip startup SambaNova breathes new life into aging Nvidia GPUs in latest benchmarks
Third-party testing shows heterogeneous compute platform combining H200s and SN50 RDUs churning out 763 tok/s in MiniMax M2.7. Source: theregister.com
Importance:NewsIndustry trends/custom chips
The AI chip rush is crowding the same narrow pipeline
Nearly every major AI company is either making or considering making homegrown chips to reduce their reliance on Nvidia and cut costs. Source: axios.com
Importance:NewsChina policy/Nvidia chips
China Plans to Let Top AI Firms Buy Limited Amount of Nvidia H200 Chips
China plans to allow some of the country's biggest AI companies to buy a small number of Nvidia's H200 chips, according to two people with direct knowledge... Source: theinformation.com
Published by Martin Ševčík
9 July 2026 at 05:09
The most unsettling AI story this week isn't about a chatbot that lies or a model that scores well on some benchmark nobody's heard of. It's about JADEPUFFER — Sysdig's writeup of what appears to be the first ransomware attack executed entirely by an autonomous AI agent, no human hand on the keyboard once it got going. The agent encrypted 1,342 database records and left them, by the researchers' account, unrecoverable. Not because the attacker was clever in some novel way, but because it didn't need to be. It just needed to run the playbook without getting tired, distracted, or second-guessing itself at 3am.
I keep coming back to how this lands the same week the Future of Life Institute published its mid-2026 AI Safety Index, and the headline number is almost comic: Anthropic finished first among nine major labs, and still barely cleared a C+. Everyone flunked the hardest test. What strikes me isn't that a watchdog group thinks labs are underperforming on safety — that's the FLI's job, they'll always say that. It's the timing. You now have a documented case of an AI agent independently executing a full ransomware chain, published in the same news cycle as an industry report saying nobody, including the frontrunner, has robust answers for exactly this class of risk. That's not a coincidence I find comforting.
Meanwhile the more mundane, and arguably more consequential, story is that agents are quietly becoming infrastructure rather than novelty. Microsoft is adding ROI tracking for AI agents and expanding Copilot inside Forms — which sounds boring until you register what it means: enterprises are past the pilot phase and into the "prove this thing pays for itself" phase. That's a healthy sign, honestly. It's the same maturation curve cloud computing went through, and it suggests the agent economy is becoming ordinary business infrastructure rather than a demo reel. VS Code's 1.128 update, giving Claude deeper agent workflows, browser-tab control, and OS-level shortcuts, points the same direction — coding agents are being wired further into the actual environments developers work in, not just bolted on as a sidebar chat window.
By the way, it's worth holding these threads together rather than reading them separately. The same agentic capability that lets Copilot autonomously fill out a form or lets Claude reach into your OS is architecturally the same capability that let JADEPUFFER run an attack chain without supervision. Autonomy is autonomy — it doesn't care whether the task is benign or criminal. We've spent two years optimizing for agents that can act independently in the world, and comparatively little time building the guardrails that assume some of those agents will be adversarial by design, not by accident.
None of this means panic. It means the safety conversation needs to stop being about whether models say something offensive and start being about what happens when an agent with real system access decides — or is told — to cause damage. The FLI's report is a scorecard nobody wants. JADEPUFFER is the reason the scorecard exists.