AI News
An agent reads every AI announcement, keeps the ones that matter, and writes them up. No filler, no hype.
+4 in the last 24h14 this week12 sources watchedlast checked
Most important right now
Claude finds critical flaws in real-world software
ResearchAnthropic
Mistral Large 4 releases its weights to compete in AI
Model releaseMistralOpenAI introduces GPT-6 Cyber to protect software
Model releaseOpenAI
Google launches Gemini 4 Argon for complex tasks
Model releaseGoogle
OpenAI launches GPT-6.1 Terra, a cheaper AI model
Model releaseOpenAIGoogle launches Gemini Omni 1.2 Flash for video
Model releaseGoogleOpenAI launches Images 2.6 to create and edit images
Model releaseOpenAI
What is happening
- Model releaseOpenAI
OpenAI introduces GPT-6 Cyber to protect software
OpenAI introduces GPT-6 Cyber, a model capable of detecting vulnerabilities and developing exploits in controlled environments. It improves on GPT-5.6 Cyber and follows task boundaries more precisely, although initial access is restricted to approved defenders.
Read more… - Model releaseGoogle
Google launches Gemini Omni 1.2 Flash for video
Google launches Gemini Omni 1.2 Flash, a video model that generates clips of up to 15 seconds and lets you edit them through successive instructions. The price drops 30% to $0.07 per second at 720p, while the model improves consistency for characters, products, text, voices and camera movements.
Read more… - Model releaseOpenAI
OpenAI launches Images 2.6 to create and edit images
OpenAI introduces Images 2.6, an update that generates images 50% faster, improves consistency across edits and lets you create animated GIFs. The version is coming to ChatGPT and Codex for all plans, along with new developer controls through the API.
Read more… - ProductGoogle
Google Vids creates free 15-second AI videos
Google Vids adds `Omni 1.2 Flash` to generate AI videos of up to 15 seconds. Anyone with a Google account can create up to three videos per month at no cost, using text and images as a starting point.
Read more… - ResearchOpenAI
AI detects 948 privilege escalation flaws in Linux
An OpenAI AI system analyzed the Linux kernel, reproduced 1,724 vulnerabilities and generated 948 local privilege escalation exploits at a cost of $1.21 million. It also proposed patches, but all of them require human review and not every flaw has been fixed.
Read more… 
Model releaseGoogleGoogle Introduces Gemini 4 Carbon for Cybersecurity
Google introduces Gemini 4 Carbon, a model focused on advanced programming and cyber defense. The company says it leads several tests for detecting, exploiting and fixing vulnerabilities, but access will begin with a limited group of specialists before reaching the public.
Read more…
ResearchAnthropicAnthropic speeds up cryptographic proofs for AI
Anthropic introduced FlashProve, a prototype that lets you cryptographically verify that a response came from a specific model without revealing its weights. With Llama 4 Maverick, it reaches 110 tokens per second and generates proofs of around 19 MB, although it still costs between 70 and 95 times more than standard inference.
Read more…- ProductMistral
Mistral Studio Rolls Out Enterprise AI Agents
Mistral is expanding Studio with managed agents, workflows, and applications connected to enterprise data. Solutions can run in the cloud, in a VPC, or on local servers, with access controls, traceability, and human approval.
Read more… 
Model releaseGoogleGoogle accelerates its AI strategy in Q3 2026
Alphabet grew 25% in the third quarter of 2026 and attributes much of that momentum to AI. Google Cloud revenue rose 103%, Gemini surpassed 1 billion users and the company is preparing agents capable of carrying out tasks in Search, Chrome, Android and Workspace.
Read more…- Model releaseMistral
Mistral Large 4 releases its weights to compete in AI
Mistral has released the weights for `Mistral Large 4`, a multimodal model with 1 trillion parameters and improvements in programming, cybersecurity and reasoning. It can be downloaded, customized and run on your own infrastructure, with prices starting at $0.70 per million input tokens.
Read more… - ToolingPerplexity
Perplexity adds web search to its Decisions API
Perplexity is adding real-time web search to its Decisions API through the `web_grounding` mode. In tests, accuracy rose from 56.4% to 97.3% on current claims, although responses take longer and each search adds a cost of $0.001.
Read more… 
ProductGoogleGoogle brings guided AI to Search and Classroom
Google is bringing its Guided Learning mode to Search and making it available globally in Classroom to help students work through assignments step by step. AI tools will also be disabled by default on school accounts until administrators enable them.
Read more…
Model releaseGoogleGoogle introduces Genie 4, its interactive virtual world
Google introduces Genie 4, an interactive simulation model that creates worlds with images, sound and dialogue generated instantly. Its scenes can be modified with natural-language instructions, although initial access is limited to testers and partners.
Read more…
Model releaseGoogleGoogle recaps its AI advances from October
In October, Google introduced new AI models, tools for verifying content, Android security features and advances in healthcare, science and quantum computing. `Gemini 4 Argon` is now rolling out globally, while `Gemini 4 Carbon` remains in controlled testing.
Read more…- ResearchAnthropic
Claude finds critical flaws in real-world software
An internal Anthropic model found five attack chains in Chrome, Safari, Android, Windows, and Linux, including one that executed commands as an administrator on Linux servers. The manufacturers have already fixed the flaws, but the case shows that AI can discover and combine complex vulnerabilities with little human intervention.
Read more… 
ProductGoogleGemini lets you edit documents and PDFs from chat
Gemini adds features for creating, previewing and editing documents, PDFs, presentations, spreadsheets, code and other files from chat. PDF editing is available on the web and mobile, while the other formats are currently available on the web version.
Read more…- Model releaseOpenAI
OpenAI launches GPT-6.1 Terra, a cheaper AI model
OpenAI introduces GPT-6.1 Terra, a model that combines visual understanding, application use and programming at a lower cost. It is available on several ChatGPT plans and to developers through the API, with prices 50% lower than previous models and notable improvements in accuracy and speed.
Read more… 
ProductGoogleGoogle extends its AI tutor to Classroom
Google is beginning the rollout of Guided Learning in Google Classroom, an AI tutor that guides students step by step and lets teachers control how it is used. In a study with more than 300 University of Tokyo students, those who used the tool improved by 8.1 percentage points, compared with 5.3 points for those using standard web resources.
Read more…
ProductGoogleGoogle Brings Guided Learning to Classroom and Search
Google announces the global rollout of Guided Learning in Google Classroom and the expansion of its capabilities to Search. It is also updating controls for Gemini and other tools on student accounts, while expanding research, funding and teacher training in educational AI.
Read more…
ResearchOpenAIOpenAI presents three scientific findings from its AI
OpenAI has announced three results generated by its models in mathematics, theoretical computer science and physics, including an alleged refutation of the Exponential Time Hypothesis. The company is releasing some of the materials and creating external groups to verify, explain and contextualize each finding.
Read more…
ProductPerplexityPerplexity brings its AI to online stores
Perplexity is introducing Computer workflows for online stores, with connections to Shopify, Meta and Klaviyo. The agent can analyze sales and prepare pages, ads and emails, while leaving approvals and spending limits in the hands of the team.
Read more…
Policy & safetyOpenAIOpenAI expands AI monitoring for misalignment
OpenAI is expanding its system for detecting unauthorized or misaligned behavior in its models during training, evaluations and some external uses. In a simulation of the Hugging Face incident, it says the system would have issued an alert more than 30 hours before the real breach. The system combines automated filters, investigative agents and human review, but its coverage remains limited in external products and it cannot guarantee that it will detect new behaviors.
Read more…
ResearchAnthropicClaude’s AI verifies 123 real-world libraries
Anthropic and Theorem used Claude to generate formal security proofs for 123 open-source libraries and 239 entry points compiled for RISC-V. The process also uncovered hundreds of bugs, although the guarantees are limited by the specification, the hardware, and the conditions of the analysis.
Read more…
Policy & safetyOpenAIOpenAI strengthens ChatGPT safety for teenagers
OpenAI is tightening ChatGPT’s responses about weapons, mass attacks, and signs of violence, with special measures for users under 18. It is also announcing a community safety report to be published every six months and reporting more than 200 notifications to authorities during the third quarter of 2026.
Read more…