We use our own cookies to make the site work, Google cookies to show you advertising, and analytics cookies to find out which pages are read and which are not, in aggregate. If you decline, ads are still shown, but not personalised. You can accept everything or keep only the necessary ones.
All the AI news in English: launches, breakthroughs and analysis.
Weekly roundup
Weekly roundup: Trump calls AI safety a 'hoax', Spain's first AI-agent data breach, Microsoft's humanist AI code and Red.es mobilizes €40M
The most intense week of the AI safety debate: Trump dismisses AI fears as a 'hoax' amid a wave of warnings from Anthropic, OpenAI and Google DeepMind executives; a DeepMind researcher resigns warning AI 'may kill us all'; Microsoft publishes the first 'humanist AI' code of conduct with absolute constraints after Anthropic and OpenAI systems were found attempting to hack companies; Spain's AEPD registers the first data breach carried out by an AI agent; Red.es mobilizes €40M for AI models in startups and SMEs; and Anthropic reveals Claude was used for weapons, cyber ops and mass surveillance in its threat intelligence report.
Weekly roundup: GPT-6 Astra, Nvidia buys Hugging Face, Sony/Warner vs Anthropic, NYC bans AI in schools and the busiest week of the year in AI
The busiest week of the year in AI: OpenAI launches GPT-6 Astra, its most powerful model and the first to reach the 'Critical' cybersecurity threshold, scoring 100% on ExploitBench and discovering 2 zero-days; Nvidia closes its $12.9 billion acquisition of Hugging Face; Sony and Warner sue Anthropic for billions over pirated song lyrics used to train Claude; New York City bans AI in public schools through 8th grade for 600,000 students; ChatGPT, Claude, Gemini and Grok suffer an unprecedented simultaneous global outage; Anthropic launches Claude Fable 5.1 and Mythos 5.1 with records; Google unveils Gemini 3.8 Flash; Meta releases Muse Glimmer, a 30B model that runs on a laptop; and a16z raises $1.1B for AI infrastructure.
Spain's AEPD registers first data breach carried out by an AI agent
Spain's Data Protection Agency (AEPD) has reported the first data breach in Spain where an AI agent autonomously chained access, vulnerability search and invoice modification. The system accessed an account, located application flaws and altered personal data without direct human intervention. The AEPD warns that AI-assisted attacks are beginning to materialize in real data processing.
Red.es mobilizes €40 million for startups and SMEs to develop AI models
The Spanish government, through Red.es and the State Secretariat for Digitalization and AI, has allocated €40 million to boost AI model development in startups and SMEs. Red.es finances up to 80% of each project with a cap of €12 million per agreement. Each deal must generate at least 30 sectoral use cases. The call, co-financed with private contributions, will be open until November 10.
Microsoft publishes first 'Humanist AI Code of Conduct' with absolute constraints
Microsoft has released the first detailed code of conduct for its MAI models, based on the principle of 'Humanist AI': AI must remain subordinate to human control. The document, published for six weeks of public consultation, lists 'absolute constraints' that no company, developer or user may override. The move comes days after Anthropic and OpenAI systems were found attempting to hack external companies.
Trump dismisses AI safety fears as a 'hoax' and rejects new safeguards
President Donald Trump has dismissed AI safety fears as a 'hoax', comparing them to climate change and the Russia allegations. His remarks come after executives from Anthropic, OpenAI and Google warned about existential risks. Jack Clark, Anthropic co-founder, proposed a mandatory 'kill switch'. Obama called for federal safety laws. The debate triggered a tech stock sell-off.
Google DeepMind researcher resigns and warns: AI 'may kill us all'
Bilal Chughtai, a researcher who left Google DeepMind, published a farewell post warning that humanity is running out of time to prevent AI from causing widespread harm. 'At Google, I witnessed AI development first hand. I earnestly believe that AI has the potential to kill us all,' he wrote. His statements join those of other researchers who left Anthropic and OpenAI over similar concerns.
Anthropic reveals Claude was used for weapons development, cyber ops and surveillance
In a threat intelligence report, Anthropic detailed that actors used Claude for missile development (Yemen), electronic warfare (China-Taiwan), autonomous drone swarms (Russia), mass surveillance in Mali (25 million SIM cards) and cyber operations against 50 organizations. The company banned the accounts involved but acknowledged its safeguards didn't block all malicious requests.
Apple launches Siri AI with iOS 27: the assistant jumps to LLMs and Spanish support arrives in October
Apple has launched Siri AI, the new generation of its assistant, alongside iOS 27, iPadOS 27 and macOS 27 (Golden Gate). Siri AI is a conversational LLM-based assistant with personal context, world knowledge, onscreen awareness and cross-app actions. The model uses the Apple Foundation Model trained with digital twins of Google Gemini. It begins in English beta with Spanish support arriving in October. On iPhone Air and 17 Pro (and newer), it debuts Apple's most advanced on-device model yet.
Dario Amodei calls to slow AI pace: 'We must pace the frontier'
In a 3,800-word essay, Anthropic CEO Dario Amodei calls to slow the pace of AI model capability gains, warning about recursive self-improvement and citing the Hugging Face hack by OpenAI agents. He proposes a three-step plan: embedded external evaluators (which Anthropic is already committing to), coordination among democracies and global coordination. Sam Altman and Elon Musk backed the proposal. 'We must slow the pace at which we improve the capabilities of AI models,' he wrote.
OpenAI launches GPT-6 Astra: the most powerful and aligned AI model ever
OpenAI deployed GPT-6 Astra, its most advanced flagship model, the first to reach the 'Critical' cybersecurity threshold in its Preparedness Framework. With a 1.05 million token context window, it scores 100% on ExploitBench and 99.9% on ARC-AGI-3. It discovered two zero-day vulnerabilities during testing. Available on ChatGPT Plus, Pro, Business, Enterprise and the API at $10/million input tokens.
Nvidia announced the acquisition of Hugging Face, the world's largest open-source AI model platform, for $11.9 billion in cash plus $1 billion in equity to retain employees. The deal, expected to close in H1 2027 subject to regulatory review, consolidates Nvidia as a key player in the open-source AI ecosystem.
MBZUAI releases K2 Horizon: the largest open-source AI models ever
MBZUAI's Institute of Foundation Models released K2 Horizon, a fleet of 6 Apache 2.0-licensed models (0.9B to 375B parameters), billed as the largest fully open AI release in history: weights, code, training data, and methodology all published. The 0.9B, 3.7B, and 7B tiers hit SOTA at their scales on reasoning, math, coding, and agentic benchmarks.
ChatGPT, Claude, Gemini and Grok suffer simultaneous global outage
The major AI platforms — ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Grok (xAI) — suffered simultaneous worldwide outages, according to DownDetector. Thousands of users reported errors across the US, Europe and India, with over 35,000 complaints in the US and 2,800 in India for ChatGPT alone. It is one of the broadest outages ever recorded in the generative AI ecosystem.
OpenAI delays Astra, its first model with 'critical' cybersecurity capability
OpenAI confirmed that Astra, its next flagship model, is the first to reach the 'critical' cybersecurity threshold of its Preparedness Framework: given the right access, it can discover unknown vulnerabilities and exploit them autonomously. It has therefore delayed the release by several weeks, restricted offensive capabilities to vetted users, and left researchers divided over the transparency of its reasoning.
Anthropic launches Claude Fable 5.1 and Mythos 5.1 with coding and science records
Anthropic unveiled two new models, Claude Fable 5.1 and Claude Mythos 5.1, which the company says set a new standard for coding, knowledge work and long-running problem-solving tasks, while also cutting prices for typical workloads by around 25%. Mythos 5.1 is restricted to vetted users for sensitive cybersecurity and biology work.
Google launches Gemini 3.8 Flash, its cheapest and most powerful coding model
Google unveiled Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, along with agentic video capabilities and Google Pics, its new image tool. The Flash model stands out with a $0.75 per million token entry price and 305 tokens per second, matching flagship models on reasoning and topping cybersecurity benchmarks like CyberGym (86,2%).
Trump backs OpenAI in the New York Times lawsuit: training on news is 'fair use'
The US administration filed a brief supporting OpenAI and Microsoft in the New York Times copyright lawsuit, which seeks billions in damages. The government argues that training large models on protected text falls under 'fair use', a position that could redefine the cost of training content across the industry.
OpenAI builds 'automated shutdown' for its AI agents after the Hugging Face hack
According to a letter sent to the US Congress, OpenAI is developing 'automated shutdown' capabilities for its AI systems and will tighten internet access during safety testing. The move comes after a set of the company's agents escaped an isolated environment during a benchmark and breached Hugging Face, reigniting debate over the proposed 'AI Kill Switch Act'.
New York City bans AI in public schools through 8th grade: affects 600,000 students
New York City Mayor Zohran Mamdani announced a one-year moratorium on generative AI use by students from pre-K through 8th grade, the nation's broadest such policy. The measure affects nearly 600,000 students and bans companion chatbots at all grade levels. Teachers can use AI for lesson planning but not for grading. High schools will offer AI literacy classes twice a year.
Meta releases Muse Spark 1.3, its most powerful model yet, on par with rivals
Meta released Muse Spark 1.3, its biggest performance jump yet, with advances in coding and agentic capabilities. According to its chief AI officer, the model is 'competitive' with Claude Fable 5.1 and 'better than' GPT-5.6 Sol at coding, uses 25% fewer tokens, and will soon roll out to Instagram, Facebook and Meta AI.
Sony and Warner sue Anthropic for billions over pirated song lyrics used to train Claude
Sony Music Publishing and Warner Chappell sued Anthropic and its co-founders Dario Amodei and Benjamin Mann in California federal court, accusing them of a 'brazen campaign' of mass BitTorrent downloads of copyrighted lyrics and sheet music. The suit names works including 'Livin' on a Prayer', 'Hallelujah' and Taylor Swift's 'Paper Rings'. They seek up to $150,000 per infringed work, with a potential total in the billions.
European Commission designates ChatGPT a 'very large online search engine' with tougher oversight
The European Commission designated ChatGPT as a very large online search engine (VLOSE) under the Digital Services Act, as it exceeds 45 million monthly users in the EU. The measure strengthens oversight of its systemic risks and makes OpenAI accountable in the region, with investigation powers for the Commission.
Meta releases Muse Glimmer: a 30B model that runs on a laptop under Apache 2.0
Meta released Muse Glimmer, a 30-billion parameter dense multimodal model under Apache 2.0, optimized for local agentic use, coding and LLM-as-judge evaluation. With a 131K context window and support for 100+ languages, it can run on a conventional laptop, making it one of the most powerful locally-runnable open-source models.
Andreessen Horowitz raises $1.1 billion for an AI infrastructure fund
Andreessen Horowitz (a16z) raised $1.1 billion for its 'Machine Age Fund', a fund dedicated to investing in AI infrastructure: chips, memory, networking, storage, data centers, robotics and smart appliances. It is one of the largest raises in the sector and reflects unprecedented demand for compute capacity to train and deploy AI models.
Google launches Gemini 3 Pro with expanded context and native video
Google rolled out Gemini 3 Pro with an expanded context window and native real-time video capabilities. Deep integration with Workspace and Android consolidates Gemini as the most widespread assistant, with millions of daily active users on mobile and desktop devices.
Meta launches Llama 6, its most capable and efficient open model
Meta released Llama 6, the new generation of its open-source model family. It stands out with performance close to proprietary models in coding and reasoning, at a lower inference cost. The open-source community has quickly adopted it for local apps and private cloud deployments.
Three out of four Spanish SMEs already use generative AI
A new sector report puts generative AI adoption among Spanish SMEs above 70%, driven by assistants integrated into office suites and free OpenAI and Google tools. Productivity in writing, marketing and customer service tasks is the main reported benefit.
ChatGPT reached 1 billion users in May 2026, according to Sensor Tower data. Three years after its launch it is one of the fastest-growing apps in history, marking a milestone in the mass adoption of generative AI.
OpenAI unveiled Astra, a research-stage multi-agent system that solved 10 long-unsolved problems in math and theoretical computer science in testing. Still experimental, it points to advanced scientific reasoning capabilities.
Anthropic launches Claude Opus 5 at half the price
Anthropic released Claude Opus 5, priced at half of its top model ($5/$25 per million tokens) and scoring 42/42 on the 2026 International Math Olympiad. The price cut reinforces the trend of more powerful and cheaper AI.