Updated May 25, 2026

AI News

A curated digest of the most important AI announcements, model releases, research, and enterprise rollouts — refreshed regularly.

Latest updates as of May 25, 2026

42 stories

SecurityTop story

Cloud Security Alliance details two-wave AI developer supply-chain attack

The Cloud Security Alliance published a May 22 analysis of TeamPCP's Shai-Hulud/Megalodon campaign against AI developer infrastructure. CSA says Mini Shai-Hulud compromised 172 npm packages and 2 PyPI packages across 404 malicious versions, then Megalodon pushed 5,718 malicious commits to 5,561 GitHub repositories in under six hours, with persistence hooks targeting tools including Claude Code and Visual Studio Code.

AgentsTop story

OpenAI Codex named a Leader in enterprise AI coding agents

OpenAI said Codex was recognized as a Leader in Gartner's 2026 Magic Quadrant for Enterprise AI Coding Agents. The company says Codex is used by more than 4 million people each week, and highlighted enterprise controls including approval gates, RBAC, customizable policies, OS-level sandboxing, auditable workspace governance, IDE and CLI surfaces, SDKs, and cloud orchestration.

OpenAI
EnterpriseTop story

Virgin Atlantic says Codex speeds refactors and app testing

OpenAI published a Virgin Atlantic case study saying the airline used Codex to ship a revamped mobile app with near-complete unit test coverage and zero P1 defects at launch. Virgin Atlantic also reported 78% to 80% codebase size reductions on some legacy refactors and said work that once took two weeks can now take about 30 minutes to an hour.

OpenAI
Enterprise

AdventHealth deploys ChatGPT for Healthcare across clinical workflows

OpenAI detailed AdventHealth's deployment of ChatGPT Enterprise and ChatGPT for Healthcare across a hospital system operating in nine states. AdventHealth says the rollout targets administrative burden, utilization-management summaries, structured rationales, and operational workflows, with an 80% reduction in time spent on some administrative tasks and an emphasis on governance and measured adoption.

OpenAI
Hardware

Hark raises $700M for a universal AI interface and hardware

TechCrunch reported that Hark, the AI lab founded by Figure AI and Archer founder Brett Adcock, raised a $700 million Series A at a $6 billion post-money valuation. Hark says it is building an agentic AI system as a universal interface for the digital world, expects to release multimodal models this summer, and plans custom hardware after that.

TechCrunch
Agents

Microsoft Foundry Labs ships new open agentic stack and benchmarks

Microsoft Foundry Labs released a May roundup with SocialReasoning-Bench for measuring whether agents act in a user's best interest, plus an open end-to-end agentic stack made up of MagenticLite, MagenticBrain, and Fara 1.5. The stack emphasizes visible reasoning, browser and local-file workflows, sandboxed code execution, human approvals for critical actions, and small computer-use models built on Qwen 3.5.

Hardware

NVIDIA Vera Rubin NVL72 and Jetson Thor win COMPUTEX AI awards

NVIDIA said its Vera Rubin NVL72 rack-scale AI supercomputer, Jetson Thor edge AI and robotics platform, and Alpamayo autonomous-vehicle platform won COMPUTEX 2026 Best Choice Awards. NVIDIA says Vera Rubin NVL72 is designed for agentic AI, reasoning, and long-context workloads, while Jetson Thor delivers up to 2,070 FP4 teraflops for physical AI and autonomous robots.

NVIDIA Blog
Research

OpenAI model disproves long-standing discrete geometry conjecture

OpenAI reported that an internal general-purpose reasoning model disproved a central conjecture in the planar unit distance problem, producing an infinite family of constructions with polynomial improvement over the long-believed square-grid bound. OpenAI says external mathematicians checked the proof and wrote companion remarks, calling the result a milestone for AI-assisted mathematics.

OpenAI
Talent

Anthropic hires Andrej Karpathy for Claude pretraining research

OpenAI cofounder and former Tesla AI director Andrej Karpathy said he is joining Anthropic. CNBC reports Karpathy will be part of Anthropic's pretraining team, building a group focused on using Claude to accelerate the research that gives the company's models their core knowledge and capabilities.

CNBC
Agents

Google brings AI agents and generative UI into Search

Google said AI Mode in Search now uses Gemini 3.5 Flash globally and introduced a redesigned AI-powered Search box. New Search agents will monitor the web in the background, send synthesized updates, help with booking tasks, and eventually generate custom interactive layouts, simulations, dashboards, and trackers with Antigravity-powered coding.

Google Blog
Security

Google expands SynthID and Content Credentials verification

Google expanded AI-content verification across Search, Gemini, Chrome, Pixel, and Google Cloud, saying SynthID has watermarked more than 100 billion images and videos and 60,000 years of audio. OpenAI, Kakao, and ElevenLabs are adopting SynthID for more AI-generated content, while a new Google Cloud AI Content Detection API is launching with trusted partners.

Google Blog
ModelsTop story

Google launches Gemini Omni Flash for multimodal video generation

Google introduced Gemini Omni, a new model family that combines Gemini reasoning with generative media, beginning with video output. The first release, Gemini Omni Flash, can use text, images, video, and audio references to generate or conversationally edit videos, is rolling out to Google AI Plus, Pro, and Ultra subscribers through Gemini and Flow, and will come to developer and enterprise APIs in the coming weeks.

Google Blog
AgentsTop story

Google previews Gemini Spark as a 24/7 personal AI agent

Google announced Gemini Spark, a cloud-based personal agent powered by Gemini 3.5 and the Antigravity harness. Spark is designed to keep working after a laptop closes, integrate with Gmail, Docs, Slides, and other connected apps, ask before high-stakes actions, and roll out first to trusted testers before a U.S. beta for Google AI Ultra subscribers.

Google Blog
ModelsTop story

Google releases Gemini 3.5 Flash for agents and coding

At Google I/O 2026, Google introduced Gemini 3.5 as a model family focused on complex agentic workflows, starting with Gemini 3.5 Flash. Google says Flash is now available globally in the Gemini app, AI Mode in Search, Antigravity, the Gemini API, AI Studio, Android Studio, and Gemini Enterprise, with claimed gains on coding and agentic benchmarks plus 4x faster output than other frontier models.

Google Blog
Security

Ocean emerges from stealth with $28M to fight AI phishing

Ocean, an agentic email-security startup founded by former Israeli cybersecurity researcher Shay Shwartz, emerged from stealth with $28 million in total funding led by Lightspeed Venture Partners. The company says AI has automated spear-phishing at much larger scale and that its small language model analyzes billions of emails each month for customers including Kayak, Kingston Technology, and Headspace.

TechCrunch
Agents

Anthropic acquires Stainless to strengthen agent connectivity

Anthropic acquired Stainless, the SDK and MCP server tooling company that has generated official Anthropic SDKs since the API's early days. Stainless creates SDKs, CLIs, and MCP servers from API specs across TypeScript, Python, Go, Java, and more, and Anthropic says the deal will help Claude agents connect more reliably to external systems.

Anthropic
Hardware

NVIDIA ships first Vera CPUs to top AI labs

NVIDIA delivered its first standalone Vera CPU systems to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure, moving the agentic-AI processor from announcement to customer evaluation. Vera packs 88 NVIDIA-designed Olympus cores, 1.2TB/s of memory bandwidth, and 50% faster per-core performance for agent sandboxes, tool calls, orchestration, and long-context retrieval workloads.

NVIDIA Blog
EnterpriseTop story

Anthropic and Gates Foundation commit $200M to beneficial AI programs

Anthropic announced a four-year, $200 million partnership with the Gates Foundation spanning Claude usage credits, technical support, and grant funding. The work targets global health, life sciences, education, and economic mobility, including public health datasets, healthcare AI benchmarks, disease-modeling support, AI tools for neglected diseases, K-12 tutoring, and agricultural productivity applications.

Anthropic
Enterprise

Khosla backs Synthetic with $10M for autonomous AI bookkeeping

Synthetic, founded by former Bench Accounting CEO Ian Crosby, raised a $10 million seed round led by Khosla Ventures to pursue a fully autonomous AI bookkeeper for accrual-based financials. The startup plans to focus on AI and software companies first, while acknowledging that current foundation models still make bookkeeping mistakes and the product remains in the design phase.

TechCrunch
Hardware

Lovable backs Atech to bring vibe coding to hardware prototypes

Danish startup Atech raised an $800,000 pre-seed round with backing from Lovable, a16z scout fund, Sequoia Scout Fund, and Nordic Makers. Atech pairs hardware starter kits with an AI chatbot that turns natural-language prototype ideas into code for working hardware builds, aiming to reduce the engineering barrier for physical products.

TechCrunch
AgentsTop story

OpenAI brings Codex to the ChatGPT mobile app

OpenAI rolled out Codex in preview on iOS and Android so users can follow active coding threads, review diffs and terminal output, approve actions, and redirect long-running agent work from a phone. The update also makes Remote SSH generally available, adds generally available Codex hooks, introduces programmatic access tokens for Business and Enterprise workspaces, and supports eligible HIPAA-compliant local Codex deployments.

OpenAI
Security

OpenAI says two employee devices were hit by TanStack supply-chain attack

After malicious TanStack package versions spread through npm, OpenAI confirmed two employee devices were affected and that a limited subset of internal source-code repositories saw unauthorized credential access. The company said it found no evidence that user data, production systems, intellectual property, or software releases were compromised and began rotating signing certificates as a precaution.

TechCrunch
Security

OpenAI updates ChatGPT to better track risk in sensitive conversations

OpenAI detailed new safety updates that help ChatGPT recognize when self-harm, suicide, or harm-to-others risk emerges over time. The system uses short-lived, narrowly scoped safety summaries for rare high-risk cases and improved safe-response performance by 50% in long suicide and self-harm evaluations, 16% in harm-to-others scenarios, and 39% to 52% across multi-conversation GPT-5.5 Instant tests.

OpenAI
Security

Twin Prime raises $10M to build frontier AI for defense and security

London-based Twin Prime landed a $10 million pre-seed round led by Expeditions to develop multimodal AI models for defense and security. The startup is building systems that reason across sensor modalities and compress perception-to-decision workflows for real-time threat response, with plans for a joint venture with European defense prime Theon.

Tech.eu
EnterpriseTop story

Anthropic launches Claude for Small Business

Announced May 13, Claude for Small Business plugs directly into QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365. It ships with 15 ready-to-run agentic workflows spanning finance, ops, sales, marketing, HR, and customer service — including automated payroll planning, month-end reconciliation, campaign management, and invoice tracking.

Anthropic
Hardware

Meta unveils four new MTIA chips for its AI data centers

Meta announced a new MTIA (Meta Training and Inference Accelerator) lineup. MTIA 300 is already deployed for training smaller ranking and recommendation models; MTIA 400, 450, and 500 are in development for generative AI inference and will launch by 2027.

AIstify
Models

NVIDIA launches Nemotron 3 Nano Omni multimodal model

Nemotron 3 Nano Omni is an open multimodal model unifying vision, audio, and language. NVIDIA reports up to 9× higher throughput than competing open models, targeting more efficient AI agents on commodity hardware.

NVIDIA Blog
Research

NVIDIA partners with David Silver's Ineffable Intelligence

NVIDIA announced a collaboration with British AI startup Ineffable Intelligence, founded by former DeepMind RL lead David Silver, to develop systems that learn through reinforcement learning rather than human data. The work will run on NVIDIA's Grace Blackwell and Vera Rubin platforms.

CNBC
Models

NVIDIA releases Star Elastic: one checkpoint, three reasoning models

NVIDIA Research introduced Star Elastic, a post-training method that embeds nested 30B, 23B, and 12B reasoning submodels inside a single checkpoint with zero-shot slicing. Operators can pick a model size at inference time without retraining.

ModelsTop story

OpenAI releases GPT-5.5 ("Spud"), its most agentic model yet

Rolled out to paid ChatGPT and Codex users on May 13, GPT-5.5 is tuned for long-running agentic tasks with minimal prompting. API access will follow once additional security guardrails are in place. OpenAI did not publish SWE-bench Verified scores, where Anthropic's Claude Mythos Preview currently leads at 93.9%.

Implicator.ai
Talent

Thinking Machines Lab loses key talent to Meta, OpenAI, and xAI

After founding employees crossed the one-year cliff and unlocked equity, Thinking Machines Lab saw a wave of departures. Meta reportedly recruited seven founding team members plus a star researcher with compensation packages worth hundreds of millions.

Research

Google DeepMind reimagines the mouse pointer with Gemini

DeepMind unveiled an AI-enabled pointer powered by Gemini that understands on-screen visual context. Users can issue shorthand commands like "Fix this" or "Show me directions" without switching windows or writing long prompts.

Google DeepMind
Agents

Google publishes patterns for long-running enterprise agents

Google's Developers Blog detailed how to build pause-and-resume agents with the Agent Development Kit (ADK). The approach uses durable memory schemas and event-driven dormancy gates — instead of stateless chatbot patterns — to support multi-week workflows like HR onboarding without losing context.

Enterprise

IBM debuts Red Hat AI Inference and OpenShift Virtualization on IBM Cloud

IBM announced two managed offerings on May 12: Red Hat AI Inference Service and Red Hat OpenShift Virtualization Service on IBM Cloud. Both are aimed at helping enterprises operationalize AI and run virtualized workloads at scale with built-in governance controls.

IBM Newsroom
Security

Microsoft's MDASH agentic security system tops CyberGym

Microsoft's new multi-model security system (codename MDASH) orchestrates 100+ specialized agents and posted an industry-leading 88.45% on the CyberGym benchmark. In the announcement, Microsoft says the system has already discovered 16 new vulnerabilities in Windows, including four critical RCE flaws.

Security

OpenAI introduces "Daybreak" cyber platform

Announced May 12, Daybreak combines OpenAI's language models with Codex's agentic capabilities to automate vulnerability detection, patch validation, and secure software development inside enterprise security workflows. The launch puts OpenAI head-to-head with Anthropic's Mythos in enterprise cyber.

Computerworld
Agents

Power Apps MCP server adds closed-loop learning for agents

Microsoft introduced closed-loop learning on the Power Apps MCP server: user corrections automatically improve enterprise agent performance using memory-based optimization and a genetic-Pareto optimization step.

Enterprise

SAP and Anthropic bring Claude to SAP Business AI Platform

At SAP Sapphire, SAP and Anthropic announced plans to embed Claude across the Business AI Platform to advance the "Autonomous Enterprise." Claude will power agentic capabilities such as financial closing, employee leave questions, and supplier order management directly inside SAP systems.

SAP News Center
Agents

SAP and NVIDIA co-define enterprise-grade agent execution

SAP and NVIDIA detailed a joint framework for secure, auditable, and governable AI agents built on NVIDIA OpenShell. The work focuses on the runtime controls enterprises need before pushing autonomous agents into production.

SAP News Center
Enterprise

SAP unveils the Autonomous Enterprise with 50+ Joule Assistants

SAP introduced a unified Business AI Platform and Autonomous Suite, deploying more than 50 domain-specific Joule Assistants across finance, supply chain, and HR. Partnerships span Anthropic, AWS, Google Cloud, Microsoft, NVIDIA, and Palantir. SAP says its Autonomous Close Assistant can compress financial closing from weeks to days.

SAP News Center
Research

Microsoft research: AI agents still struggle with long workflows

A Microsoft study using the new DELEGATE-52 benchmark tested frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) across 52 professional workflows. The team found models lose ~25% of document content over 20 interactions on average, with severe corruption in 80% of conditions. Only Python programming hit "ready" status at 98%+ accuracy.

The Register
Enterprise

OpenAI launches the "OpenAI Deployment Company"

A new entity dedicated to helping organizations build and deploy AI for mission-critical work. The Deployment Company starts with $4B in initial backing from 19 global investment firms and consultancies, and absorbs Tomoro to bring on roughly 150 Forward Deployed Engineers and Deployment Specialists.

OpenAI