Anthropic said Claude outputs will carry machine-readable marks after it signed the EU AI Act Article 50(2) Code of Practice on AI-generated content transparency, The Hindu reported. New EU-facing Claude models from August 2 must support the marks at launch, with older models on a transition path. Text gets an imperceptible watermark that survives copy, paste, and light edits, while supported image files can carry signed provenance across Claude apps and major clouds. Why it matters: provenance is becoming a product default under European rules. Caveat: marks can miss after heavy edits and do not prove Claude was the sole author.
University of Illinois researchers report ESMDynamic in Nature Communications, predicting residue contact dynamics from sequence instead of a static fold. Built on ESMFold and trained on structure ensembles plus molecular dynamics, it outputs dynamic contact probabilities, occupancy, and coarse contact kinetics across temperatures. On mdCATH and ATLAS it matched or beat ensemble generators such as AlphaFlow at far lower compute, with maps released for more than 18,000 human proteins. Why it matters: dynamics often decide function and drugability. Caveat: sequence-only forecasts still need wet-lab or simulation checks before design use.
Korea University and Catholic University of Korea teams introduce AUGUST in npj Digital Medicine, a sequential question-answering system for gastric whole-slide diagnosis. It generates the next clinical question from prior findings, embeds the slide in that context, and answers only along valid hierarchical pathways. On 23,072 slides spanning 19 dependent tasks and nearly 80,000 question-answer pairs, it beat multiple-instance learning, pathology vision-language models, and foundation baselines. Why it matters: pathology AI fails when it ignores conditional reasoning. Caveat: gastric-focused training leaves broader organ generalization unproven here.
Chinese teams at Huazhong University of Science and Technology and Sichuan University report XunZi in Nature Biomedical Engineering, trained on 24.4 million papers and hundreds of terabytes of multi-omics across 21,008 genes and 5,850 diseases. It proposes therapeutic targets with testable mechanisms, and in Parkinson's models flagged aberrant CHK2 and IRAK4 activity. Chk2 inhibition rescued dopaminergic neuron loss and motor deficits in mice, with further demos in non-small-cell lung cancer. Why it matters: hypothesis generation is becoming wet-labable. Caveat: mouse rescue is not a therapy, and independent replication is still required.
DeepSeek is inviting open-source developers into beta tests for DeepSeek Harness, software meant to turn language models into multi-step agents, the South China Morning Post reported. Harness lead Cui Tianyi, hired in March from quant trading, posted the call as the firm pushes agent tooling beside its cheap V4 Flash model. Agentic harnesses became a competitive front after Anthropic's Claude Code showed demand for coded workflows. Why it matters: Hangzhou labs are contesting the agent control plane, not only token prices. Caveat: no release date was given, and the harness group remains short-staffed.
Chairmen of the House Select Committee on China and the Homeland Security Committee asked DoorDash CEO Tony Xu for documents on Chinese AI evaluation and deployment, CNBC reported. The letter cites co-founder posts describing Moonshot's Kimi K2.6 in lower-level code-review workloads, continuing a probe that already reached Cursor and Airbnb. DoorDash said it supports American AI leadership and will discuss safe use of frontier and open-weight models. Why it matters: Washington is treating PRC open-weight adoption as a security process issue. Caveat: a document request is scrutiny, not a ban or finding of misuse.
Onyx Security announced a $113 million Series B led by Bessemer Venture Partners, with Cyberstarts, TCV, Conviction, FirstMark, Vintage, QuantumLight, and G Squared, taking total funding to $153 million. The New York and Tel Aviv firm sells a control plane that discovers agents and inspects actions before they execute across browser, endpoint, SaaS, and cloud surfaces. It reports more than 1.1 million agents secured and 66 million sessions inspected after a fourfold revenue jump since stealth. Why it matters: agent adoption is outrunning governance. Caveat: session counts are self-reported and depend on where traffic is routed.
Bengaluru's Enlife Research closed a Rs 6 crore seed led by Piper Serica for an AI-read blood biomarker platform aimed at Alzheimer's, The Economic Times reported. Founded by IISc neuroscientist Deepak Kumaran Nair with Jojo Jacob and Anish Mathew, it wants a routine blood draw that flags risk years earlier than PET, MRI, or CSF tests built on Western cohorts. Funds will validate assays and push the first test toward clinical-grade use with IISc, NIMHANS, and Centre for Brain Research partners. Why it matters: India-priced neurodegeneration diagnostics are scarce. Caveat: the firm is pre-revenue at TRL 3-4.
Indian AI startups are commanding seed rounds of $3 million to $5 million and Series A cheques above $10 million, investors told The Economic Times, versus about $2 million to $4 million and $5 million to $8 million for non-AI peers. That premium is forcing smaller and micro funds to change playbooks if they still want meaningful ownership rather than token stakes. Why it matters: early-stage AI in India is repricing into capital-heavy bets. Caveat: interview averages can overstate the market, and thin wrappers will still struggle to raise.
In the revised 2026-27 budget, Tamil Nadu Finance Minister N. Marie Wilson said the state will establish Arivagam as India's first dedicated AI and innovation city, The Hindu reported. Plans sketch an AI university, skill centre, semiconductor hardware testing lab, multi-domain incubators, and housing, with Rs 5 crore for a feasibility study. The same speech floated Guidance 3.0, an AI-powered single-window investor service. Why it matters: a major state is packaging AI as industrial policy. Caveat: the capital line funds a study, so delivery and site selection remain unproven.
YC-backed Discovered Materials closed a $9 million seed led by Lightspeed India Partners, with Peak XV and angels including Paul Graham and Gokul Rajaram, TechCrunch reported. Founders Advaith Sridhar and Akash Ramdas run agent swarms that propose semiconductor materials, then check them with physics models aimed at heat in AI chips. They released hundreds of candidate materials and a Material Discovery Bench, and plan to patent useful routes for licensing to chipmakers. Indian VC leadership gives the round a clear India capital signal. Caveat: no AI-discovered chip material from this startup class has yet shipped at commercial scale, and synthesis remains the bottleneck.
Caixin reported that ByteDance CEO Liang Rubo told a companywide meeting the firm's large language models have fallen further behind leading overseas systems, even as Doubao stays competitive and Seedance sits at the video frontier. Liang said ByteDance will keep developing its own foundation models rather than chase short-term rankings with outside models, accepting a temporary lag to harden the long-term stack. The stance matches founder Zhang Yiming's delayed-gratification line against distillation shortcuts. For China's model race, it is a rare admission from inside a top lab. Caveat: the account relies on people familiar with a closed meeting, not an official transcript.