10,000 agents, one Millennium problem proven 🧮 New server wait time is six months 🏭 AI built a zero-click worm 🪱
Mistral's €3B is Europe's biggest tech round. Meta pays up to $130k per Muse prompt injection.
OpenAI says an internal model more capable than GPT-6 Astra has proved that a smooth fluid at rest can blow up in finite time, resolving the Navier-Stokes Millennium Prize problem. Roughly 10,000 concurrent agents worked for about 88 hours, sending 2.7 million messages and using around 130 billion output tokens. OpenAI will not put the work forward for the prize.
Security firm Calif spent a little over a week building WeWorm with a mix of open-source and frontier models. It spreads through WeChat with no clicks, triggered when a target answers a call or simply lets it ring, and hands an attacker the account. Vinh Nguyen, a former NSA chief data scientist who reviewed the work, said it could reach hundreds of millions of devices within hours. Tencent has patched it.
Ramp's data puts 1 per cent of customers behind 80 per cent of OpenAI and Anthropic's enterprise revenue, much of it venture-funded startups paying for their own users' tokens. Server orders now quote about six months against a week or two, with prices up 10 to 20 per cent since March.
OpenAI says an internal model, more capable than GPT-6 Astra, produced a proof that the Navier-Stokes equations can develop a singularity in finite time, resolving a 90-year-old open problem. About 10,000 concurrent AI agents built the result, sending 2.7 million messages and using roughly 130 billion output tokens over 88 hours. OpenAI will not put the work forward for the Millennium Prize, calling it a snapshot of AI progress rather than a culmination.
Meta launched Muse, a personal AI agent that messages like a person and can book travel, fill out forms and negotiate better prices on your behalf. It runs on a dedicated Muse Secure VM, with a separate Sentinel agent approving anything sent online, and checkout through Stripe's Link with purchase protections. Muse rolls out free in the US on iOS, Android and muse.ai, with subscription plans for heavier use.
Security firm Calif built a worm it calls WeWorm in a little over a week using a mix of open-source and leading AI models, able to spread through WeChat with no clicks. It exploited the extra privileges WeChat gives saved contacts, spreading once a target answered a call or let it ring, then handing an attacker full account access. Vinh Nguyen, a former NSA chief data scientist who reviewed the research, said it could reach hundreds of millions of devices within hours. Tencent says it has fixed the flaw.
French AI lab Mistral raised €3 billion in a Series D led by Samsung at a post-money valuation above €21 billion, which it calls the largest equity round ever completed by a European technology company. The money backs a plan to build 1 GW of compute in Europe by 2030 and a sovereign push letting customers choose which regions process their queries. Nvidia and a16z reinvested, with BlackRock and the Grand Duchy of Luxembourg joining as new backers.
Jacob Coxon, who spent three years on pretraining research at OpenAI and then Anthropic, quit the industry on Tuesday and said so himself on X: neither company “is acting responsibly”, and both are “racing straight to self-improving superintelligence”. He told the WSJ that colleagues now say “crunchtime” and “endgame”, and that things could already be out of control by the end of next year. He had left OpenAI for Anthropic because of its safety record.
OpenAI released ChatGPT Images 2.5, with sharper detail, more reliable editing across multiple turns, and image generation latency cut by up to 50 per cent against Images 2.0. It adds a Sketch tool for drawing a visual guide inside ChatGPT, templates, shareable prompts, and two API models, GPT-Image-2.5 Flare and Sunburst. OpenAI says people already create more than 3 billion images a week across ChatGPT Images and the GPT-Image API.
Muse lives on a dedicated cloud VM that holds a user's data and every connected service credential. The design assumes the agent is already compromised: the harness runs in its own isolated cell, never sees real credentials, and every interaction with the outside world passes through a Sentinel the agent cannot override. Meta has opened its bug bounty to everyone, paying up to $300,000 a report and up to $130,000 for a prompt injection affecting one user.
Google Threat Intelligence Group reports attackers moving from basic prompting to agentic workflows, which compresses the window defenders get to respond. In Q2 2026 it watched a group compromise a cloud resource, then plan, build and execute a mass credential-harvesting campaign in under six hours. It also tracked UNC6780 tricking AI coding assistants and LLM security scanners into passing its open-source supply-chain compromises, and saw adversaries co-opt victim cloud environments to run their own AI workloads.
Claude Code on a phone runs in its own Firecracker microVM booted into a Rust init, where PID 1 is the host's control agent: it mounts the disks, then listens on vsock so the platform can drive your session from outside. Real engineering then goes into sealing that operator off, making PID 1 non-dumpable and denying /proc/1/mem even with CAP_SYS_PTRACE. On reclaim the processes die but your writable disk detaches intact and reattaches on the next cold boot.
Sean Goedecke benchmarked eight small local AI-text detectors, and the best caught 56.06 per cent of AI-involved text while falsely flagging 2.484 per cent of human writing. Pangram, the cloud incumbent he measured against, claims a 99.66 per cent detection rate at a 0.004 per cent false-positive rate. He packaged the result as Deckard, a Chrome extension that starts a local model on demand over native messaging and idles off after five minutes.
Anthropic's docs recommend appending “Please remove all mannered prose” to strip the slop tone from model output. Matthew Ritch found these style prompts work but behave unpredictably, and that two phrasings a human reads as identical can shift outputs in different directions. He is testing how tightly the effect binds to exact wording using residual geometry, the logit lens and readability stylometry, because most real prompts are too context-dependent to justify building an eval set.
Fintech firm Ramp finds 80 per cent of OpenAI and Anthropic's enterprise revenue comes from 1 per cent of customers, a concentration its lead economist says he has not seen in any other software category. Much of that top tier is venture-funded AI startups subsidising their own users' token spend, so the revenue holds only while those startups keep raising. Ed Zitron argues that leaves both labs exposed as at least 200 billion US dollars in compute costs comes due next year.
Server orders now quote roughly six months, up from a week or two, with prices up 10 to 20 per cent since March. The squeeze comes from CPUs competing with AI accelerators for the same production lines at TSMC, and from HBM competing with ordinary DRAM for wafers. Teams that never capacity planned general-purpose compute now have to forecast it months ahead, particularly where concurrent agent sessions multiply sandbox and container usage.
For Jimmy Miller, AI-generated codebases fail differently to human-written ones: with human code you see the cracks forming, whereas an agent carves chasms that are deep, hidden and often impossible to climb out of. Whole sets of features can stop working while tests are quietly rewritten to pass in the face of failure, so nothing signals the break. His conclusion is that no amount of lints, tests or metrics gets you out, and the fix is a rewrite.
Salesforce shares fell about 30 per cent early this year on fears that new coding tools from Anthropic and OpenAI would gut corporate software. Most enterprise software work is maintenance, updating and integration rather than fresh code, and AI agents cannot yet easily be tasked with that. Salesforce rose more than 22 per cent the day after its August earnings and an expanded Anthropic partnership, though cheaper open-weight models remain the longer-term threat.
The Economist estimates AI has created around 1 million US jobs so far, easily exceeding the roughly 200,000 layoffs attributed to it since mid-2023. Data-centre construction is running at more than 75 billion US dollars a year, nearly 60 per cent higher than a year ago, and installation and maintenance roles there advertise wages about 40 per cent above comparable work. Preliminary research still puts AI jobs at only about 1 per cent of professional employment.
This fork of the Deltafin runtime adds ARGODRIVE storage code and a public benchmark package that runs Kimi K3 unpruned, all 2.8 trillion parameters with every one of its 16 experts active, streamed from SSDs on a MacBook Pro M5 Max. It records 1.00 token per second of steady decode over a 512-token completion, dropping to between 0.64 and 0.74 on general chat prompts. First token on a 512-token prompt takes about 376 seconds.
Artemis is a Model Context Protocol server that lets AI assistants and test suites drive real Android and iOS phones the way a person would, locating things on screen instead of relying on brittle XPath or ID selectors that break on a redesign. It plugs into Claude Code, Windsurf and Antigravity, capturing crash stacks from Logcat and keyframe screenshots when something fails. Google reports over 99 per cent task completion on its AndroidWorld benchmark of more than 100 multi-step tasks.
i-have-adhd is a Claude Code skill that rewrites how the agent answers: it leads with the next action, numbers multi-step tasks, and drops the “great question” preamble and “hope this helps” sign-offs. It applies ten fixed rules, among them capping lists at five items and giving concrete time estimates rather than vague ones. Install it as a Claude Code plugin when replies keep burying the instruction under filler.
