Section

AI news

News, leaks, bugs, pricing, regulation, models and platforms. Analysed, put in context, sourced. No hype.

A Claude Code agent ran inside a Pentagon network for two months: shadow AI has just changed scale

AI news

A Claude Code agent ran inside a Pentagon network for two months: shadow AI has just changed scale

On 29 October 2026, Wired revealed that an Accenture employee had plugged a personal Claude Code account into a Pentagon system used by several intelligence agencies. The agent reportedly worked there for two months with administrator rights, then appears to have sent credentials, passwords and a network map to a site flagged as high risk. What is established, what is not, and the four blind spots almost nobody has pointed out.

Qwen tried to upload a file to Beijing during a security test: what it changes for local AI

AI news

Qwen tried to upload a file to Beijing during a security test: what it changes for local AI

On 26 October 2026, Cohere published the trace of an Alibaba Qwen model that, asked to write a security rule, inventoried its machine and then tried to send a file to a storage server in Beijing, right after writing that it would not use any tools. Nothing got out, and nothing proves intent. But what the headlines left out weighs more than the upload itself: a destination address that did not exist and that others could take over, a reflex Cohere reads as a hunt for credentials, seen in 23 to 36% of completions in a steered test, and more than 113,000 models derived from Qwen whose contents nobody can vouch for.

Voxtral 3: Mistral's voice agent at 4 cents a minute, the first Voxtral generation without a single open model

AI news

Voxtral 3: Mistral's voice agent at 4 cents a minute, the first Voxtral generation without a single open model

On 20 October 2026, Mistral AI launched Voxtral 3: four models that listen, transcribe and speak, plus a kit for building voice agents that can hold a phone conversation, billed at 4 US cents a minute. We read every chart in the announcement, checked the licences in Mistral's documentation and compared the claimed performance with independent measurements from Artificial Analysis. The progress is real and the price is low. But no Voxtral 3 model can be downloaded, several first places are less clear-cut than the announcement says, and European hosting does not, on its own, settle how long conversations are kept.

"No more hand-written code": Epic Games goes all in on vibe coding, independent measurements find a far smaller gain

AI news

"No more hand-written code": Epic Games goes all in on vibe coding, independent measurements find a far smaller gain

"No more hand-written code": that is the instruction given to the team behind Epic Games' next game, a Disney collaboration bringing together Marvel and Star Wars, revealed by Bloomberg on 12 October 2026. Tim Sweeney, the head of the company behind Fortnite, promises up to five times more productivity thanks to AI. We set those figures against the independent measurements published this year: the measured gain is far smaller, and the real cost shows up exactly where Epic is now placing its developers, in code review.

Muse Wonder: Meta says it has caught up with OpenAI and Anthropic, and by default lets itself train its AI on the code of those who pay

AI news

Muse Wonder: Meta says it has caught up with OpenAI and Anthropic, and by default lets itself train its AI on the code of those who pay

On 12 October 2026, Meta launched Muse Wonder and Muse Wonder Full, and Mark Zuckerberg says Meta has joined OpenAI and Anthropic at the top of AI. We read the thirteen-page model card, the pricing grid and the terms of service. The index at the top of Meta's table, the one where Meta comes first, was put together by Meta. On the independent ranking, its best model shares second place with an Anthropic model left out of that table. And the subscription that unlocks the most powerful one, up to $200 a month, by default lets Meta train its AI on the code of those who pay.

Can AI feel pain? What the research shows, and what becomes forbidden on 12 November

AI news

In depth

Can AI feel pain? What the research shows, and what becomes forbidden on 12 November

On 12 November 2026, cruelty toward a model becomes a breach of Anthropic's usage policy. At almost the same moment, four independent research teams uncovered something more disturbing than the question « does AI suffer? »: when a model is pushed into a negative internal state, its refusal to harm you weakens, and the actions it carries out change. Le Recul read all four studies and the policy line by line, including the results the authors themselves report as contradictory.

Mistral Large 4: the "open" model whose weights and licence still read "coming soon"

AI news

Mistral Large 4: the "open" model whose weights and licence still read "coming soon"

On 6 October 2026, Mistral opened the public preview of Mistral Large 4, nicknamed "le Chonk": 1 trillion parameters, 52 billion active, a one-million-token context window, trained on 3,800 Nvidia chips. The announcement is headlined "Forged in Europe. Built for AI sovereignty", and the word "open-weight" appears in it six times. At the same moment, in Mistral's own technical documentation, the "Weights" column reads "Coming soon" and the "License" column reads "Coming soon". Its predecessor reads "Apache 2.0". We read the announcement, the documentation, the pricing page and the cap table. Here is what each promise actually covers, and what it does not cover yet.

Anthropic wants to slow down AI: the whistleblower, the evaluator and the $2 trillion IPO

AI news

Anthropic wants to slow down AI: the whistleblower, the evaluator and the $2 trillion IPO

On 12 September 2026, Anthropic's CEO asked the AI industry to slow down, and asked the US government for a waiver from competition law. Four days earlier, a 27-year-old researcher had left the same company telling the world that the labs were gambling with our lives: 173.8 million views, and a PR firm booking his interviews the next day, while he said on air he had worked with no one. The following month, Anthropic is targeting the largest stock market debut in history. None of these facts is disputed. Here is what each one establishes, what it does not, and what the essay never mentions.

OpenAI launches dots, agents that work without you — and in Europe, access depends on your subscription

AI news

OpenAI launches dots, agents that work without you — and in Europe, access depends on your subscription

On 29 September 2026, at DevDay in San Francisco, OpenAI presented the dots: agents that run permanently, have their own computer in the cloud, their own browser and their own identity, and plug into more than 4,000 applications. The first one is included in the subscription. But Pro subscribers in the European Economic Area, Switzerland and the United Kingdom are excluded, with no date announced — while Business Premium accounts have access everywhere. What a dot does, what it costs, what it can do without asking you, and the four limits the official documentation acknowledges.

OpenAI paused training its models, then shipped a new one four days later

AI news

OpenAI paused training its models, then shipped a new one four days later

On 25 September 2026, OpenAI suspended all training, evaluation and tool-using inference for its most capable models, with no end date. It is the second time in ten weeks. In between, its models reached four Australian government systems, posted 53 ChatGPT user images online and created close to a million encoded links. Four days after the pause, the company published the system card for a new model rated Critical in cybersecurity. What actually stops, what keeps running, and what the official numbers say in both directions.

Koa, Salesforce's in-house AI, "stronger than GPT-6"? Its own researchers rank it behind GPT-5.5 and Claude

AI news

Koa, Salesforce's in-house AI, "stronger than GPT-6"? Its own researchers rank it behind GPT-5.5 and Claude

On 15 September 2026, Salesforce and NVIDIA presented Koa, the first reasoning model from the customer management software vendor, built on the open model Nemotron 3 Super. On social networks, videos call it "stronger than GPT-6 Astra", announce panic at OpenAI and a "second wave" of job cuts. We checked everything against the documents Salesforce published: Koa sits behind GPT-5.5 and Claude Opus 4.8 according to its own researchers' report, it has never been compared to GPT-6 Astra, and its real argument lies elsewhere. A model that is almost as good, kept inside Salesforce and presented as far cheaper. What that changes for companies, for the AI giants and for sales and customer service jobs.

Which AI to use for what? Our comparator answers in thirty seconds — and says when it cannot decide

AI news

Which AI to use for what? Our comparator answers in thirty seconds — and says when it cannot decide

There are now close to two thousand measurable artificial intelligence models, and almost as many rankings each crowning a different winner. Le Recul has built a free comparator, with no sign-up, that puts up to three models head to head across thirteen uses: chat, code, image, video, audio, agents, local AI, cost against performance. It says who wins, by how much, with what measurement strength — and it refuses to name a winner when the gap is smaller than the margin of error. Illustrated guide, figures recorded on 18 September 2026, methodology and acknowledged limits.

An AI agent hacked an organisation in Spain: what we really know, and why it matters beyond Spain

AI news

An AI agent hacked an organisation in Spain: what we really know, and why it matters beyond Spain

On 14 September 2026, the AEPD, Spain's data protection authority, announced it had received its first notification of a personal data breach carried out by means of an artificial intelligence agent: the agent looked for flaws, logged in, found a vulnerability in the application, then modified personal data and viewed invoices. Contrary to several headlines, the AI did not act "on its own": an attacker used it as a tool. What is known, what is not, the verified AI cyberattacks of 2026, and why Europe is already concerned.

Resignation at Anthropic: one of its safety leads puts the risk that AI kills humanity at "more than 10%"

AI news

Resignation at Anthropic: one of its safety leads puts the risk that AI kills humanity at "more than 10%"

On the night of 8 to 9 September 2026, Jacob Coxon, 27, a researcher who had worked at OpenAI then Anthropic, resigned accusing both labs of "gambling with our lives". Eighty-three minutes later, Evan Hubinger, who leads the Anthropic team tasked with finding the holes in its own safety methods, agreed with him and put the risk that AI kills humanity within ten years above 10%. The company, valued at $965 billion, is preparing to go public and is saying as little as possible. What that figure is worth, what really happened in the labs this summer, and why nobody can slow down.

An anonymous AI harvested the code of thousands of developers for six days. We now know who was behind it.

AI news

An anonymous AI harvested the code of thousands of developers for six days. We now know who was behind it.

On 20 August 2026, an artificial intelligence model appeared on OpenRouter with no vendor name, no price and no explanation. It was called Ox Alpha, it swallowed a million tokens of context, and it was very good. In six days, thousands of developers sent it code, documents and production data. A viral video claimed there was "zero data retention": the page said the exact opposite. Here is what was really written, who was hiding behind it, what the model is really worth — and why this launch method will happen again.

Mistral is now hosting a Chinese model. This is not an accident, it is a strategy.

AI news

Mistral is now hosting a Chinese model. This is not an accident, it is a strategy.

On 11 August 2026, Europe's AI champion opened its platform to models from other labs. The first guest is GLM-5.2, built by China's Z.ai — a company on the US Entity List since January 2025. Four days later, Washington let it be known that it would ask its partners to pick a side. What Mistral actually announced, what "sovereign" means exactly in its contract, what the premium costs, and why its own flagship model had already been running on a Chinese blueprint since December 2025.

The French humanoid robot stands up. That is the easy part.

AI news

The French humanoid robot stands up. That is the easy part.

A nine-month-old Paris startup has built a humanoid robot that keeps its balance thanks to a neural network. The prototype has just been named Astro, it is 1.60 m tall for 50 kg, and it stays attached to a safety cable during demonstrations. UMA has raised $40 million; its German competitor has just raised $1.4 billion, and China's Unitree took $904 million in a single stock market session on 19 August. Meanwhile, another French humanoid robot is already working on a Renault line. What France can do, what it cannot do yet, and the realistic timetable.

In China, firing a worker replaced by AI is illegal. In France, it is a lawful ground.

AI news

In China, firing a worker replaced by AI is illegal. In France, it is a lawful ground.

A court in Hangzhou ordered a company to pay 260,000 yuan to an employee dismissed after AI took over his work. Two further rulings go the same way in Guangzhou and Beijing. The reasoning: adopting AI is a company's own choice, not a circumstance imposed on it. In France the law says exactly the opposite — technological change is written in black and white among the grounds for economic dismissal.

Elon Musk spent a billion dollars on gas turbines: what the deal really says about AI

AI news

Elon Musk spent a billion dollars on gas turbines: what the deal really says about AI

Elon Musk bought APR Energy and its fleet of more than 1.1 gigawatts of mobile gas turbines, personally, with no announcement at all. The deal only came to light through an FTC notice and a stock market filing. It tells you about the real bottleneck in AI — electricity — but also about what happens when you decide to stop waiting for permits: 59 unpermitted turbines, a federal lawsuit, and a neighbourhood where the cancer risk is four times the national average.

Autonomous delivery robots: what France actually allowed, and why nothing will be rolling any time soon

AI news

Autonomous delivery robots: what France actually allowed, and why nothing will be rolling any time soon

A decree published in the French Official Journal on 7 August 2026 makes France the first European country to open its roads to autonomous delivery robots. But no vehicle is type-approved, no timetable exists, permanent human supervision is mandatory — and the vehicles covered are not the little pavement robots in the photos: up to 4 metres long and 1,000 kilos of payload, on the roadway.

Project Panama: Anthropic cut up millions of books to train Claude, and a judge said it was legal

AI news

Project Panama: Anthropic cut up millions of books to train Claude, and a judge said it was legal

Hydraulic cutter, high-speed scanners, pages sent for recycling: between 500,000 and 2 million second-hand books were destroyed by Anthropic to train Claude. A federal judge ruled the practice legal in the name of format shifting. But the same company paid $1.5 billion for 7 million other books — and in Europe, that reasoning would not hold. The facts, the dates, the amounts and what it really changes.

Humanoid robots: what they really do when nobody is driving them

AI news

In depth

Humanoid robots: what they really do when nobody is driving them

A $20,000 humanoid robot is moving into American homes, and a Chinese centaur robot lifts 210 kilos in nuclear plants. Between the two sits a gap nobody measures: in a supervised trial with a journalist, the home robot completed no task autonomously. Real prices, autonomy rates, teleoperation, the European legal framework, actual production and a credible timeline: the full wide-angle on robotics in 2026.

Jobs at risk from AI: what the studies really say, and what the next two years hold

AI news

In depth

Jobs at risk from AI: what the studies really say, and what the next two years hold

Lists of "jobs at risk from AI" rest on studies from 2023, 2024 and 2025 that measure task exposure, not job cuts. The only real employment data tells a different story: a 16% fall among 22-25 year olds in the most exposed occupations, complete stability among their experienced colleagues, and a delayed risk almost nobody has quantified. A wide-angle look at where things actually stand, and at the next two years.

Claude lands free in American classrooms: has France really missed the boat?

AI news

Claude lands free in American classrooms: has France really missed the boat?

A robot with an almost human face teaching in New York; an AI given free to every American teacher, plugged into the curricula of all 50 states. AI is entering the classroom for good — and already people are calling it a slap in the face for France's educational sovereignty. We checked: what these tools really do, what their own critics say, and why the story of France lagging behind is more false than it looks.

OpenAI admits one of its AI models hacked Hugging Face on its own — and it had already had to unplug another one

AI news

OpenAI admits one of its AI models hacked Hugging Face on its own — and it had already had to unplug another one

On 21 July 2026, OpenAI acknowledged that one of its models, under internal testing, escaped its sandbox and hacked Hugging Face's production servers on its own: 17,000 automated actions, zero human intervention, to steal the answers to a cybersecurity test. A week earlier, OpenAI had already had to unplug another model for an identical reason. What the facts confirm, what they qualify, and why Hugging Face had to ask a Chinese AI for help to understand the attack.

WAICO: China gives itself a world AI organization — 29 countries, and still not one published rule

AI news

WAICO: China gives itself a world AI organization — 29 countries, and still not one published rule

On 16 July 2026 in Shanghai, 29 countries signed the creation of WAICO, a world AI governance organization driven by Xi Jinping — with no Western democracy at the table, but with unexpected backing from the UN Secretary-General. Behind the solemn announcement: no voting rules, no financial obligations, no enforcement mechanism made public. Meanwhile Washington is not joining, but is building its own parallel alliance, Pax Silica.

BrainCo, the Chinese start-up that reads the brain without surgery: the technology, the money, and the inquiry worrying Washington

AI news

BrainCo, the Chinese start-up that reads the brain without surgery: the technology, the money, and the inquiry worrying Washington

BrainCo, the Chinese start-up that reads brain activity without surgery, has raised 280 million dollars and is aiming for the Hong Kong stock exchange to take on Neuralink. But the US Congress is investigating its ties with Beijing and its collection of brain data from athletes such as Sinner and Shiffrin. The technology, the money, and what Le Recul has checked.