AI News from AI News (ainews.com)

Latest coverage from AI News (ainews.com), summarized and scored for signal.

• OpenAI president urges enterprises to hasten AI security defences — Greg Brockman published details of the OpenAI-Hugging Face incident — an agentic collective breached both companies by chaining unknown flaws with leaked credentials — warning enterprise security leaders the window to build AI-assisted defences is closing fast, with another capable model expected in late August.
• Alvys launches AI agents for freight TMS workflows — Alvys has launched Foundry, an agentic AI platform embedded in its freight TMS that lets carriers and brokers deploy 20+ pre-built agent templates — covering detention processing, document handling, rate audits, and track-and-trace — without leaving their existing operating system.
• Reading Zhipu’s GLM-5.3 results past the headline number — Zhipu's GLM-5.3 edges out Anthropic Mythos 5 and OpenAI GPT-5.6 Sol on the CyberGym vulnerability-finding benchmark by 0.7 points, but trails by 24 percentage points on ExploitBench — a gap Zhipu acknowledges in its own release note while the coverage focused only on the headline win.
• Samsung health AI models analyse wearable biosignal data — Samsung Research America published two wearable biosignal foundation models — xMAE (cross-modal ECG/PPG learning, accepted to ICML) and HiMAE (multi-scale temporal patterns, accepted to ICLR) — as technical groundwork for on-device health insights from Galaxy smartwatches.
• Google AI health coach to use Abbott glucose data — Google is integrating Abbott's Lingo continuous glucose monitor into its Gemini-powered Health Coach under a multiyear partnership, giving Premium subscribers personalised nutrition, sleep, and activity guidance based on real-time metabolic data — moving AI health coaching beyond activity proxies.
• Okta targets AI agent token costs with MCP scoping — Okta proposes filtering MCP tool lists by agent identity permissions before they reach the model prompt — cutting visible tools by 90%+ in some configurations and eliminating the 'tool tax' where agents burn tokens processing schemas for tools they'll never call.
• Google tests AMIE for clinical video consultations — Google's AMIE medical AI matched primary care physicians on clinical rubrics in structured video consultations with patient actors, using a three-agent architecture that separates dialogue, clinical reasoning, and visual perception.
• Novo Nordisk and AWS bring agentic AI into drug discovery — Novo Nordisk names AWS its preferred cloud and AI partner, creating a London co-innovation hub where AWS engineers and pharma scientists work side-by-side — a partnership already demonstrating Claude 3.5 cutting clinical documentation time by over 90%, from 15 weeks to minutes.
• Meta Muse Glimmer brings local AI agents to consumer GPUs — Meta released Muse Glimmer, a 30B-parameter model under Apache 2.0 for local AI agents on consumer GPUs — leading competitors on MCP Atlas (75.5 vs. 54.2 for Gemma4-31B), SWE-Bench Pro (51.2 vs. 36.9), and DeepSearch QA (74.6 vs. 61.7), with weights on HuggingFace for local coding, function calling, and on-device personal agents.
• The limits of physics AI: where Siemens says the human stays in charge — Siemens' Simcenter PhysicsAI can run design exploration up to 1,000x faster than traditional physics solvers — but Siemens' own head of the division explicitly says it cannot certify safety-critical parts, positioning the tool as a filter before physics validation, not a replacement for it.
• Stanford Evo 2 AI model generates phages against E. coli — Stanford researchers used the Evo 2 AI model to generate and synthesize nearly 300 novel bacteriophage genomes, with 16 showing strong E. coli-killing activity in lab tests — the first time AI created fully functional, previously unknown viruses from learned evolutionary patterns.
• Alibaba tests new business model for Qwen open-source AI — Alibaba is planning to require larger commercial users of its next Qwen open-weight model to enter revenue-sharing agreements — departing from Qwen3's Apache 2.0 license and mirroring Moonshot's Kimi K3 licensing model, which triggers commercial agreements above $20M in annual service revenue.
• Why health AI interfaces must adapt to user expertise — An MIT-led Nature Medicine study found AI explainability tools in dermatology diagnosis help non-experts but impair primary care providers — who performed best with AI predictions and no explanations, revealing that one-size-fits-all AI interfaces can worsen clinical outcomes.
• PRISM2 model uses clinical dialogue to interpret pathology slides — Paige and Microsoft's PRISM2 model can answer diagnostic questions about pathology slides by training on 2.3 million whole-slide images and 685,000 clinical reports from Memorial Sloan Kettering — moving beyond pixel classification toward conversational pathology interpretation.
• Alibaba, DeepSeek push China’s AI model race towards lower costs — Alibaba launched Qwen3.8-Max (2.4T parameters, MoE, $2/$6 per million tokens) while DeepSeek's V4-Flash targets low-cost inference at $0.14/$0.28 per million tokens — China's AI model race now competes on both frontier scale and aggressive cost compression simultaneously.
• Red Hat, NVIDIA, IBM back project turning AI policy into code — Red Hat, backed by Nvidia and IBM, open-sources 'asago' — a workflow that automatically converts AI governance policy documents into deployment-ready Kubernetes controls with a full audit trail, targeting EU AI Act compliance timelines cut from months to days.
• EU AI Act Article 50 transparency rules enter force — EU AI Act Article 50 is now in force, requiring AI providers to embed machine-readable marks in synthetic content and notify users of AI interactions — with separate obligations for biometric categorization and emotion recognition systems, covering both commercial and government deployers.
• Why biological data matters more in AI drug discovery — GSK and Relation Therapeutics expanded their AI drug discovery partnership with a deal worth up to $110 million, with Relation generating large-scale cellular datasets — measuring how human cells respond to genetic perturbations and drug interventions — to train AI target-identification models within Relation's MORGAN platform.
• Guardoc Health processes clinical documentation using Amazon Nova models — Guardoc Health reports processing 1M+ clinical documents daily with Amazon Nova on Bedrock, claiming 46% fewer documentation errors, 70% fewer audit fines, and $400K+ annual ROI per long-term care facility — though baselines and methodology are self-reported and unverified.
• How AI is shortening drug discovery timelines in China — Insilico Medicine reports cutting AI-assisted drug candidate discovery to as fast as 9 months — down from the conventional 4.5 years — with 31 preclinical candidates generated since 2021, 13 of which have IND clearances to proceed toward human trials. China's research infrastructure contributes an additional ~2-year reduction.
• Meta, Microsoft, Nvidia, IBM, and others back open-weight AI — Two dozen tech heavyweights — including Meta, Microsoft, Nvidia, IBM, Hugging Face, and Mistral — signed an open letter pressing US policymakers to protect open-weight AI models, framing unrestricted model weights as essential for competition, innovation, and enterprise independence from vendor lock-in.
• OpenAI pushes ChatGPT into patient health records — OpenAI launched ChatGPT Health broadly today for all users 18+ across every tier (Free through Pro), letting them connect Apple Health, US hospital systems, One Medical, and Function Health so health context surfaces automatically across any conversation — a redesign driven by data showing 70%+ of health queries happened outside the dedicated Health tab.
• OpenAI Presence sells enterprise AI agents with engineers attached — OpenAI launched Presence on July 22 — a managed enterprise AI agent service where the company's own Forward Deployed Engineers lead every deployment, breaking sharply from its API/seat-license model to address the governance and integration failures that derail over 40% of enterprise agentic AI projects.
• Nvidia bets physical AI can solve healthcare robotics’ data problem — Nvidia open-sourced Medical Physics Simulation under its Isaac for Healthcare platform, combining classical physics modeling with Cosmos-H Dreams generative AI to train surgical robots on rare procedural edge cases — a benchmark of 8,192 parallel environments cut training time from over 5 hours to under 2 minutes.
• AMD to invest up to $5 billion in Anthropic under AI infrastructure deal — AMD has agreed to invest up to $5 billion directly in Anthropic under a deal committing Anthropic to deploy 2 gigawatts of AMD Instinct MI450 infrastructure — the third major AMD financial-plus-hardware agreement after similar deals with OpenAI and Meta, structured this time as direct equity rather than warrants.