Latest AI News

Anthropic's Project Glasswing, using its Claude Mythos Preview model, has found over 10,000 high- or critical-severity vulnerabilities across critical software in just one month, working with roughly 50 partners. Cloudflare alone found 2,000 bugs with a false positive rate better than human testers. Mozilla found 271 vulnerabilities in Firefox 150, ten times more than with Claude Opus 4.6. Anthropic also scanned 1,000+ open-source projects, finding 6,202 high- or critical-severity flaws, though patching remains the key bottleneck.

Microsoft has launched an AI-guided image search experience on Bing, available now for U.S. users on desktop and mobile web with no sign-in required. Accessed via a New Version toggle in the Images section, the opt-in feature replaces traditional grid results with AI-organized categories, labeled sections, and short summaries explaining each group. The update aims to reduce visual overload and help users move from browsing to deciding faster, with cited sources included. Microsoft plans to expand the feature to additional markets soon.

Perplexity has open-sourced Bumblebee, a read-only security scanner built to protect developer machines from supply-chain threats. Used internally to secure systems behind Perplexity, Comet, and Computer, Bumblebee scans for risky packages, editor extensions, browser extensions, and AI agent configs like MCP. It supports three profiles: baseline, project, and deep. Crucially, it never invokes package managers or runs install scripts, preventing scanners from triggering the attacks they detect. Available as an open-source Go project for macOS and Linux.

Microsoft is rolling out Office app updates next week that let users move the floating Copilot button out of documents and back to the ribbon. The button, which appeared in Word, Excel, and PowerPoint, angered many Excel users because it obstructs cells and couldn't be fully disabled. Users can right-click the "Dynamic Action Button" to reposition it. Partner group product manager Katie Kivett acknowledged user frustration, and the change follows Microsoft removing Copilot buttons from Windows 11 apps last month.

Waymo has suspended freeway driving across all US markets due to concerns about construction zones, while also pausing service in Atlanta and San Antonio over flooded road issues. The moves follow a software recall after robotaxis were caught driving through flooded Texas roads at elevated speeds. Freeway trips were previously available in San Francisco, Los Angeles, Phoenix, and Miami. The Alphabet-owned company currently handles about 500,000 rides per week and is targeting 1 million paid rides weekly.

Leonardo AI has integrated 3D object generation directly into its platform, allowing creators to go from 2D concept images to production-ready 3D models without switching tools. Users generate an initial image, create multiple reference views using the 3D Reference View Creator blueprint, then export models as .glb files compatible with rigging, game engines, and 3D printers. The feature supports quad and triangle mesh types, PBR and shaded materials, and targets use cases including indie game development, e-commerce visualization, and physical fabrication.

Ideogram has launched an MCP (Model Context Protocol) integration that turns AI assistants like Claude and ChatGPT into design agents. The tool allows users to generate images, create designs, and train custom models directly within chat interfaces, without switching to a separate app. Beyond Claude and ChatGPT, Ideogram MCP is also compatible with Cursor, Hermes, and other platforms, streamlining AI-powered design workflows inside existing tools.

OpenAI has been named a Leader in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents, recognizing its Codex platform used by over 4 million people weekly and enterprise clients including Cisco, Datadog, Dell Technologies, and NVIDIA. Gartner highlighted Codex's strengths in agentic development, enterprise governance, sandboxing, and flexible deployment. Recent upgrades include GPT-5.5 integration, HIPAA compliance, and mobile support. Cisco used Codex to build its AI Defense security platform, cutting delivery time from several quarters to weeks.

Spotify and Universal Music Group have signed landmark licensing agreements enabling a new generative AI tool that lets fans create covers and remixes of songs from participating artists. Launching as a paid add-on for Spotify Premium users, the tool is built around consent, credit, and compensation, giving artists and songwriters a direct share of revenue generated. Spotify Co-CEO Alex Norström and UMG Chairman Sir Lucian Grainge both emphasized the artist-centric, responsible AI approach driving the initiative.

Spotify announced Personal Podcasts at its 2026 Investor Day, a new feature letting Premium users generate short, private AI-audio episodes using text prompts, PDFs, or links. Drawing on world knowledge and each user's Spotify taste profile, the tool can create daily briefings, topic deep dives, or weekly roundups. It rolls out to U.S. Premium users next month with monthly credits included. Spotify also launched an interactive Q&A feature for podcasts, now live for Premium mobile users in the U.S., Sweden, and Ireland.

Spotify has partnered with Universal Music Group to let Premium subscribers create AI-generated covers and remixes of licensed songs, with participating artists receiving a revenue share. The tool will launch as a paid add-on, though pricing and a release date were not disclosed. Spotify co-CEO Alex Norström emphasized consent, credit, and compensation as core principles. The deal contrasts with rivals like Suno and Udio, which faced major label lawsuits, including Suno settling a $500 million suit with Warner Music Group.

OpenAI has released Appshots, a new feature in Codex for Mac, designed to capture app context and make it easier to integrate Codex into everyday Mac workflows. The feature was announced by Ari Weinstein, who described it as enabling seamless integration of Codex across Mac applications. Appshots allows Codex to understand what a user is doing on their Mac, broadening how the AI coding tool can be applied across different tasks and apps.

Runway has launched Aleph 2.0, a video editing tool now available in its new Edit Studio on the web. The update allows users to edit a single frame, preview the change, and then have Aleph 2.0 automatically propagate that edit consistently across the entire video clip. This eliminates the tedious process of applying manual frame-by-frame adjustments, streamlining workflows for video creators who need cohesive, consistent edits applied quickly across full clips.

Anthropic has launched 28 new security and compliance integrations for Claude, powered by the new Claude Compliance API. The API gives enterprise IT and security teams programmatic access to conversation content from Claude Enterprise — including chats, files, and projects — and activity events across Claude Enterprise and the Claude Platform. Partners include Cloudflare, CrowdStrike, Microsoft Purview, Okta, Palo Alto Networks, Datadog, and Zscaler, spanning DLP, SIEM, eDiscovery, identity, and AI security posture management categories.

Cursor CEO Michael Truell has announced a limited-time promotion doubling usage limits for new users invited to the platform's Teams plan. The offer applies to any new users brought onto an existing Cursor Teams plan over the next 30 days, with those invitees receiving double the standard usage allocation throughout their entire first month. Current Teams plan members can take advantage of the deal by inviting colleagues during the promotional window.

Krea has launched LoRAs for Krea 2 in beta, introducing its most powerful fine-tuning system to date. The new feature lets users train the Krea 2 image model on their own specific style, object, or character with high precision. This gives creators significantly deeper control over generated outputs compared to previous tools, making Krea 2 a more flexible and consistent option for personalized image generation workflows and creative projects.

Microsoft Research has released MagenticLite, the next generation of Magentic-UI, an experimental agentic application designed to run efficiently on small models. It works across browsers and local file systems in a single workflow and is powered by two purpose-built models: MagenticBrain, a 14B-parameter orchestration model fine-tuned from Qwen 3 14B, and Fara1.5, a computer-use model family available in 4B, 9B, and 27B sizes that nearly doubles Fara-7B's web navigation performance and sets new state-of-the-art results among small computer-use models.

An OpenAI reasoning model has autonomously disproved a nearly year-old conjecture in discrete geometry known as the planar unit distance problem, first posed by Paul Erdős in 1946. The model constructed an infinite family of point configurations yielding polynomially more unit-distance pairs than previously known, disproving the longstanding belief that square grid constructions were optimal. The proof, verified by external mathematicians including Fields medalist Tim Gowers, uses sophisticated tools from algebraic number theory, marking the first time AI has autonomously solved a prominent open mathematical problem.

Stability AI has released Stable Audio 3.0, a family of four music and audio models trained on fully licensed data. Three models — Small SFX, Small, and Medium — are open-weight and available on Hugging Face, while the Large model is API-only for enterprise use. Key features include variable-length generation up to 6 minutes 20 seconds, full song composition on portable devices, LoRA fine-tuning support, and audio inpainting. Users own outputs and can commercialize them under the Stability AI Community License.

OpenAI is strengthening content provenance for AI-generated images through three measures: becoming a C2PA conformant generator, partnering with Google DeepMind to embed invisible SynthID watermarks in images from ChatGPT, Codex, and the OpenAI API, and launching a public verification tool at openai.com/verify. The tool checks uploaded images for both Content Credentials and SynthID signals to confirm whether they were OpenAI-generated. If no signals are detected, the tool avoids definitive conclusions, as provenance data can be stripped.

Anthropic has added two enterprise features to Claude Managed Agents: self-hosted sandboxes and MCP tunnels. Self-hosted sandboxes let tool execution run on your own infrastructure or with providers like Cloudflare, Daytona, Modal, or Vercel, while Anthropic handles the agent loop. MCP tunnels allow agents to reach private MCP servers inside corporate networks without exposing them publicly, using a lightweight gateway with end-to-end encryption. Self-hosted sandboxes are in public beta; MCP tunnels are in research preview requiring access requests.

SpaceX has made its IPO filing public, revealing plans to list on Nasdaq under ticker SPCX in what could be the largest IPO ever, targeting a $1.75 trillion valuation and raising around $75 billion. Elon Musk will serve as CEO, CTO, and board chairman. Starlink generated over $11 billion in 2025, more than half of total revenue. SpaceX directed 60% of 2025 capital spending, roughly $20 billion, to its AI division housing Grok, which lost billions despite only 22% revenue growth.

Meta laid off roughly 7,800 employees — about 10% of its 78,person workforce — on Wednesday as part of a restructuring tied to AI investment. An email obtained by Business Insider told affected workers their roles were eliminated to help offset other company investments. Laid-off staff receive 16 weeks of severance plus two weeks per year of service, 18 months of COBRA coverage, and three months of job search assistance through Lee Hecht Harrison. Meta also plans to shift over 7,000 employees to new AI initiatives.

Granola has launched AI Meeting Briefs, a feature that automatically prepares users for upcoming calls as they join. Available now on Mac, Windows, and iPhone, Briefs generate a two-to-three bullet summary covering who you're meeting, past discussions, and current priorities. Granola prepares these overnight by researching past notes, recent Gmail threads if granted access, and web sources. Every fact is cited, and users can review the steps Granola took to build each brief. No manual prompting is required.

OpenAI has launched Guaranteed Capacity, a new offering that gives enterprise customers reserved access to OpenAI compute for critical products, AI agents, and customer workflows. Organizations can choose one- to three-year commitments, with discounts scaling based on annual spend levels. Capacity can be drawn down across OpenAI's full product portfolio and supported cloud providers. The program targets businesses that need predictable infrastructure for production systems and long-term AI adoption plans without risking compute shortages as demand grows.

Google unveiled the Universal Cart at I/O 2026, an AI-powered shopping hub built on Gemini models that works across Search, Gemini, YouTube, and Gmail. It consolidates items from multiple merchants, proactively flags product incompatibilities, tracks price drops, and leverages Google Wallet for payment perks. Checkout via Google Pay is supported at Nike, Sephora, Target, Walmart, and Shopify merchants. Google also expanded its Universal Commerce Protocol to Canada, Australia, and the UK, and introduced the Agent Payments Protocol for secure agentic purchases.

Google is expanding its AI content transparency tools across Search, Gemini, Chrome, Pixel, and Cloud. SynthID, its digital watermarking technology, has now watermarked over 100 billion images and videos and 60,000 years of audio. SynthID verification in the Gemini app has already been used 50 million times globally and is expanding to Search and Chrome. Google is also launching an AI Content Detection API on Google Cloud, and partners including OpenAI, ElevenLabs, and Kakao are adopting SynthID for their own AI-generated content.

Google announced major AI upgrades to Search at I/O 2026, including Gemini 3.5 Flash as the new default model in AI Mode, which has surpassed one billion monthly users. A redesigned Search box — the biggest upgrade in 25 years — supports text, images, files, videos, and Chrome tabs as inputs. New Search agents will monitor the web 24/7 for personalized updates, while agentic booking and coding features can build custom dashboards and mini apps on the fly.

Google unveiled Gemini 3.5 at Google I/O, a new model family combining frontier intelligence with agentic capabilities. The first release, Gemini 3.5 Flash, outperforms Gemini 3.1 Pro on coding and agentic benchmarks including Terminal-Bench 2.1 at 76.2% and MCP Atlas at 83.6%, while running four times faster than other frontier models. It is available now via the Gemini app, Google Search AI Mode, and the Gemini API. Gemini 3.5 Pro is being used internally and will roll out next month.

Google has introduced Gemini Omni Flash, the first model in its new Omni family, capable of creating and editing videos from any combination of inputs including images, audio, video, and text. Announced by Google DeepMind CTO Koray Kavukcuoglu, the model supports conversational, multi-turn video editing while maintaining scene consistency. It is rolling out to the Gemini app, Google Flow, and YouTube Shorts, with image and audio output modalities planned for future support.

Google has launched Antigravity 2.0, a standalone desktop application for macOS, Linux, and Windows that replaces the IDE-centric Agent Manager with a fully agent-optimized experience powered by the latest Gemini models. New features include dynamic subagents, asynchronous task management, JSON hooks, scheduled tasks via cron, and new slash commands like /goal and /grill-me. The app is available to enterprises and expands beyond coding into general knowledge work, with a CLI, SDK, and API also launching alongside it.

Google I/O 2026 kicked off on May 19th at the Shoreline Amphitheatre in Mountain View, with a keynote livestreamed at 10AM PT on Google's YouTube channel. Gemini is front and center, with expected updates to AI models, agentic features, and Search. Google previewed Android 17 AI features including AI-generated widgets at a separate Android Show last week. Smart glasses from Samsung and Gucci are also rumored. Hardware surprises are unlikely, as Google already announced the Googlebook laptop and Fitbit Air.

Andrej Karpathy, former Tesla AI director and OpenAI co-founder, has joined Anthropic to return to frontier large language model research and development. Karpathy described the next few years at the LLM frontier as especially formative and expressed excitement about rejoining an R&D environment. He also noted he remains passionate about education and plans to resume his educational work in the future, suggesting his teaching efforts will continue alongside his new role at Anthropic.

OpenAI is strengthening content provenance by becoming a C2PA Conforming Generator Product, partnering with Google DeepMind to embed SynthID invisible watermarks in images generated via ChatGPT, Codex, and the OpenAI API, and launching a public verification tool at openai.com/verify. The tool checks uploaded images for both C2PA metadata and SynthID watermarks to confirm whether content originated from OpenAI. The dual-layer approach addresses metadata's vulnerability to being stripped through file transformations or screenshots.

Google has launched Gemini for Science, a collection of AI tools and experiments available on Google Labs designed to accelerate the scientific method. Three prototype tools include Hypothesis Generation built with Co-Scientist, Computational Discovery built with AlphaEvolve and ERA, and Literature Insights built with NotebookLM. Google is also launching Science Skills, integrating over 30 life science databases including UniProt and AlphaFold. Enterprise partners including BASF, Klarna, Daiichi Sankyo, and Bayer Crop Science are already using the tools in private preview.

Google announced real-time collaborative design updates to Stitch at I/O 2026. The Stitch Agent now streams work directly to the canvas, letting users design, steer iterations, and reflow ideas via text or voice prompts. Users can import existing codebases and design files, generate shareable links through Google AI Studio, export screens to Google Antigravity for backend integration, or publish directly to the web via Netlify. The updates are available globally starting today.

Apple has previewed a suite of Apple Intelligence-powered accessibility updates coming later this year, including enhanced VoiceOver and Magnifier with detailed image descriptions, natural language Voice Control navigation, and an improved Accessibility Reader with translation support. New on-device generated subtitles will automatically caption uncaptioned videos across iPhone, iPad, Mac, Apple TV, and Vision Pro. Apple Vision Pro also gains eye-tracking-based power wheelchair control, launching with Tolt and LUCI drive systems in the U.S.

Google Labs has updated Pomelli, its AI tool for small and medium-sized businesses, with new agentic capabilities. The Pomelli Agent helps users build a Business DNA brand identity by uploading product docs and photos or chatting with the agent. Two new features include Brand Books, which generate comprehensive guides with custom images, fonts, and colors, and a website builder that designs and launches a complete site in a few clicks. Millions of product shots and social campaigns have already been created on the platform.

Microsoft has announced new Surface Pro for Business and Surface Laptop for Business devices powered by Intel Core Ultra Series 3 processors, available now in select markets. The Surface Laptop for Business starts at $1,299.99 and the Surface Pro for Business at $1,949.99. Key features include up to 35% better graphics than MacBook Air M5, up to 23 hours battery life, a new integrated privacy screen, Secured-core PC security, and on-device AI processing. Snapdragon X2 models with 80% faster local AI inferencing arrive later this year.

SpaceX has filed its S-1 prospectus with the SEC for what could be the largest IPO in history, set to list on Nasdaq under ticker SPCX. The company reported $18.67 billion in revenue in 2025, with Starlink satellite internet accounting for over $11 billion. SpaceX lost $4.9 billion last year while spending $20.7 billion in capital expenditures. Elon Musk's supervoting shares give him 85 percent control. The company, which now includes xAI, targets a $28.5 trillion total addressable market.

Amazon has launched Alexa Podcasts, a new feature for Alexa+ subscribers that generates custom AI audio episodes on demand. Users simply tell Alexa a topic, review an outline, and adjust length or direction conversationally before Alexa creates a recording with AI-generated host voices. Content draws from over 200 news publications, including the Associated Press, Reuters, and the Washington Post. Episodes are delivered via Echo Show or the Alexa app and the feature is currently available to U.S. Alexa+ customers.

A jury unanimously dismissed all claims in Musk v. Altman after roughly two hours of deliberation, finding that Elon Musk's breach of charitable trust claim against OpenAI CEO Sam Altman and president Greg Brockman was barred by the statute of limitations. A related claim that Microsoft aided and abetted the breach also failed, as did a restitution claim. US District Judge Yvonne Gonzalez Rogers accepted the advisory verdict. Musk announced on X he plans to appeal, calling the ruling a calendar technicality rather than a decision on the merits.

OpenAI and Dell Technologies are partnering to bring Codex to hybrid and on-premises enterprise environments via the Dell AI Data Platform and Dell AI Factory. Codex, one of OpenAI's fastest-growing products with over 4 million weekly developers, will connect to governed enterprise data stored on Dell infrastructure. The collaboration also covers ChatGPT Enterprise and API-based solutions, enabling customers to build, test, and deploy AI agents securely within their own infrastructure for software development and broader knowledge work.

Anthropic has acquired Stainless, a company founded in 2022 that specializes in SDK and MCP server tooling. Stainless has powered every official Anthropic SDK since the Claude API launched and is used by hundreds of companies to generate SDKs, CLIs, and MCP servers across languages including TypeScript, Python, Go, Java, and Kotlin. The acquisition aims to expand Claude's agent connectivity, allowing AI agents to reach more data sources and tools as Anthropic shifts focus from models that answer to agents that act.

Cursor has launched Composer 2.5, a significant upgrade to its AI coding agent built on Moonshot's Kimi K2.5 open-source checkpoint. Key improvements include scaled training with 25x more synthetic tasks than Composer 2, a new targeted textual feedback technique for more precise reinforcement learning, and better effort calibration and communication style. Priced at $0.50/M input and $2.50/M output tokens, Cursor is also training a larger model with SpaceXAI using 10x more compute on Colossus 2.

Pope Leo XIV will publish his first encyclical, Magnifica Humanitas: On the Protection of Human Dignity in the Age of Artificial Intelligence, on May 25 at the Vatican's Synod Hall. Signed on May 15 — the 135th anniversary of Leo XIII's landmark labor encyclical Rerum Novarum — the document is expected to provide moral guidance on AI and the digital revolution. Speakers at the presentation include Anthropic co-founder Christopher Olah and Cardinal Víctor Manuel Fernández.

Boston Dynamics' humanoid robot Atlas has demonstrated the ability to lift and carry a mini-fridge using AI-driven behaviors. The robot coordinates its entire body to manage heavy objects, balancing complex contact points with accuracy and reliability. The demonstration highlights Atlas's growing physical capability and whole-body coordination, showcasing how AI enables the robot to handle real-world manipulation tasks that require managing multiple simultaneous contact points under load.