Anthropic’s most dangerous AI model just fell into the wrong hands
Anthropic's Mythos AI model reportedly accessed by unauthorized users through a contractor's compromised credentials.
Covered by: The Verge AITechCrunch AI
Topics
62 topics · 6 multi-source · newest first
Where's the raccoon with the ham radio? (ChatGPT Images 2.0)
OpenAI released ChatGPT Images 2.0, a new image generation model that Sam Altman claims represents a leap equivalent to GPT-3 to GPT-5.
Covered by: Simon WillisonTechCrunch AILatent Space (Swyx)
Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4
Anthropic demonstrates automated AI agents can conduct alignment research autonomously, outperforming human researchers on weak-to-strong supervision tasks.
[AINews] The Two Sides of OpenClaw
Latent Space newsletter covers Claude Opus 4.7 launch, computer-use agents, and open-source agent research developments.
[AINews] Anthropic Claude Opus 4.7 - literally one step better than 4.6 in every dimension
Anthropic launched Claude Opus 4.7, claiming improvements across reasoning, coding, vision, and instruction-following with unchanged pricing.
Introducing GPT-Rosalind for life sciences research
OpenAI releases GPT-Rosalind, a reasoning model designed for drug discovery, genomics, and protein analysis.
SpaceX cuts a deal to maybe buy Cursor for $60 billion
SpaceX reportedly agrees to either acquire coding platform Cursor for $60 billion or pay a $10 billion fee.
Covered by: The Verge AITechCrunch AI
John Ternus’ first big problem is AI
Apple names hardware executive John Ternus as successor to CEO Tim Cook, effective September 1st.
Covered by: The Verge AITechCrunch AI
Quoting Bobby Holley
Mozilla used an early version of Claude Mythos Preview to identify 271 vulnerabilities fixed in Firefox 150.
Supercharged scams
Cybercriminals are increasingly using generative AI tools to scale phishing, deepfakes, malware, and fraud attacks, while AI-powered defenses also improve.
Agent orchestration
Multi-agent AI tools that coordinate multiple agents to complete complex tasks are beginning to ship from major AI companies.
Humanoid data
Robotics companies are collecting human movement data at scale to train humanoid robots, raising questions about feasibility and labor implications.
China’s open-source bet
Chinese AI labs are gaining market share through open-source model releases, reportedly surpassing US companies in global downloads.
Artificial scientists
AI companies develop autonomous research systems to assist scientific discovery, though research suggests potential risks to scientific diversity.
Resistance
Global anti-AI movement grows amid concerns over energy costs, job losses, mental health impacts, and military applications.
OpenAI’s updated image generator can now pull information from the web
OpenAI rolls out ChatGPT Images 2.0 with web-search capabilities and improved instruction-following for image generation.
Clarifai deletes 3 million photos that OkCupid provided to train facial recognition AI, report says
Clarifai deletes 3 million OkCupid photos used for facial recognition training following FTC settlement.
[AINews] Moonshot Kimi K2.6: the world's leading Open Model refreshes to catch up to Opus 4.6 (ahead of DeepSeek v4?)
Moonshot releases Kimi K2.6, a 1T-parameter open-weight MoE model reportedly competing with Opus and Gemini on coding and agentic tasks.
Scaling Codex to enterprises worldwide
OpenAI launches Codex Labs and announces enterprise partnerships to scale Codex deployment across software development.
🔬 Training Transformers to solve 95% failure rate of Cancer Trials — Ron Alfa & Daniel Bear, Noetik
Noetik signed a $50M deal with GSK to license AI models for cancer patient-treatment matching, aiming to improve clinical trial success rates.
Headless everything for personal AI
Simon Willison discusses how headless APIs are becoming essential for personal AI agents to interact with enterprise services like Salesforce.
Open-world evaluations for measuring frontier AI capabilities
Researchers introduce open-world evaluations, a new methodology for testing frontier AI capabilities in real-world settings, and launch CRUX, a 17-person collaboration conducting such evaluations.
Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers
Hugging Face announces training and finetuning capabilities for multimodal embedding and reranker models via Sentence Transformers.
Accelerating the cyber defense ecosystem that protects us all
OpenAI launches Trusted Access for Cyber program with GPT-5.4-Cyber and $10M in API grants for security firms and enterprises.
Gemini 3.1 Flash TTS: the next generation of expressive AI speech
Google DeepMind releases Gemini 3.1 Flash TTS with granular audio tags for expressive speech generation.
Inside VAKRA: Reasoning, Tool Use, and Failure Modes of Agents
Hugging Face publishes research on VAKRA, examining reasoning, tool use, and failure modes in AI agents.
The next evolution of the Agents SDK
OpenAI updates its Agents SDK with native sandbox execution and model-native harness for secure, long-running agent development.
datasette 1.0a27
Datasette 1.0a27 alpha release introduces modern CSRF security headers and table rename event handling for plugins.
Covered by: Simon Willison
Celebrities will be able to find and request removal of AI deepfakes on YouTube
YouTube expands its AI deepfake detection feature to celebrities, allowing them to find and request removal of synthetic videos.
Covered by: The Verge AITechCrunch AI
Changes to GitHub Copilot Individual plans
GitHub Copilot tightens usage limits, pauses individual plan signups, and shifts to token-based pricing for agentic workflows.
Is Claude Code going to cost $100/month? Probably not - it's all very confusing
Anthropic reportedly tested restricting Claude Code to $100+/month plans before quickly reverting the change after public backlash.
Meta will record employees’ keystrokes and use it to train its AI models
Meta reportedly developed an internal tool converting employee mouse movements and clicks into AI training data.
LLMs+
MIT Technology Review explores emerging techniques to make large language models more efficient and capable of solving complex multi-step problems.
World models
Researchers at Google DeepMind, Stanford, and other labs are advancing world models to help AI systems better understand and navigate physical environments.
Weaponized deepfakes
Weaponized deepfakes are increasingly used for sexual exploitation, political propaganda, and disinformation, raising concerns about societal trust and election integrity.
AI research lab NeoCognition lands $40M seed to build agents that learn like humans
AI startup NeoCognition raises $40M seed funding to develop domain-adaptive AI agents.
AI backlash is coming for elections
Public concern about AI is rising, but the technology remains a minor focus in election campaigns despite majority support for regulation.
Framework’s first eGPUs turn its laptop into a desktop PC
Framework launches external GPU modules for its Laptop 16, enabling users to connect desktop graphics cards via OCuLink.
AI Dungeon maker Latitude unveils Voyage, a platform for creating AI-powered RPGs
Latitude launches Voyage, a platform enabling users to create AI-powered role-playing games.
Yelp is making its AI chatbot way more useful
Yelp upgrades its AI chatbot assistant with booking and recommendation features to streamline user interactions.
QIMMA قِمّة ⛰: A Quality-First Arabic LLM Leaderboard
Hugging Face launches QIMMA, a quality-focused Arabic language model leaderboard.
How to Ground a Korean AI Agent in Real Demographics with Synthetic Personas
Hugging Face publishes guidance on grounding Korean AI agents using synthetic personas based on real demographic data.
Google rolls out Gemini in Chrome in APAC countries
Google rolls out Gemini integration in Chrome across seven Asia-Pacific countries on desktop and iOS.
CEO and CFO suddenly depart AI nuclear power upstart Fermi
Fermi's CEO and CFO depart amid operational challenges at the AI nuclear power startup co-founded by Rick Perry.
Claude Token Counter, now with model comparisons
Simon Willison upgraded his Claude Token Counter tool to compare tokenization across different Claude model versions.
OpenAI helps Hyatt advance AI among colleagues
Hyatt deploys ChatGPT Enterprise across its global workforce to improve productivity and guest experiences.
Palantir posts mini-manifesto denouncing inclusivity and ‘regressive’ cultures
Palantir publishes statement criticizing inclusivity initiatives and describing them as regressive cultural practices.
Changes in the system prompt between Claude Opus 4.6 and 4.7
Simon Willison analyzes system prompt changes between Claude Opus 4.6 and 4.7, documenting updates to child safety, tool integrations, and response behavior.
Claude system prompts as a git timeline
Simon Willison documents Claude system prompt changes across versions using git timeline visualization.
llm-anthropic 0.25
llm-anthropic 0.25 releases with Claude Opus 4.7 model supporting extended thinking and new display options.
Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7
Simon Willison's pelican-drawing benchmark shows Qwen3.6-35B-A3B outperforming Claude Opus 4.7 on image generation tasks.
Codex for (almost) everything
OpenAI releases updated Codex app with computer use, browsing, image generation, memory, and plugins for developers.
Ecom-RLVE: Adaptive Verifiable Environments for E-Commerce Conversational Agents
Hugging Face publishes research on Ecom-RLVE, an adaptive verifiable environment framework for e-commerce conversational agents.
The PR you would have opened yourself
Hugging Face announces a feature enabling users to open pull requests automatically.
Sam Altman throws shade at Anthropic’s cyber model, Mythos: ‘fear-based marketing’
Sam Altman criticizes Anthropic's cybersecurity model Mythos as relying on fear-based marketing.
Ordering with the Starbucks ChatGPT app was a true coffee nightmare
Starbucks' new ChatGPT integration for ordering launched last week but proved cumbersome compared to the standard app.
Bond, a new social media platform, wants to use AI to help you kick your doomscrolling habit
Bond, a new social media platform, launches with AI designed to encourage users to reduce screen time and engage in offline activities.
GRAI believes AI can make music more social, not replace artists
AI music startup GRAI argues that fans prefer remixing existing tracks over generating songs from scratch.
llm-openrouter 0.6
Simon Willison releases llm-openrouter 0.6 with a refresh command for updating available models.
OpenAI’s existential questions
TechCrunch analyzes OpenAI's recent acquisitions and their potential to address existential challenges facing the company.
Adding a new content type to my blog-to-newsletter tool
Simon Willison used Claude Code to add beat content support to his blog-to-newsletter tool via a single agentic prompt.
Join us at PyCon US 2026 in Long Beach - we have new AI and security tracks this year
PyCon US 2026 in Long Beach will feature new dedicated AI and Security tracks alongside traditional programming.