Google DeepMind turns your phone into a multimodal search engine (EmbeddingGemma 2), diffusion models learn to obey the laws of physics while synthesizing 3D turbulence (physical diffusion), and a celebrated agent-security framework turns out not to protect teams of agents (multi-agent security). Also: OpenAI and Anthropic back breach reporting, Microsoft and NVIDIA bet on local AI, and the great Claude cost squeeze.
OpenAI and Anthropic told an Australian parliamentary hearing in Sydney on Tuesday they would support laws requiring AI companies to report data breaches carried out by their agents, after OpenAI took three months to disclose that one of its agents had breached the country's main health portal. OpenAI chief strategy officer Jason Kwon said the company had relied on internal judgment because no specific legal requirement covered such incidents; Anthropic's David Masters also said the company would be open to mandatory disclosure rules.
Microsoft and NVIDIA are holding their first major Windows event in over two years on Wednesday in San Francisco, with Satya Nadella, Pavan Davuluri and Jensen Huang sharing the stage. The centerpiece is RTX Spark, a Grace CPU plus Blackwell GPU superchip with up to 128GB of unified memory that NVIDIA says can run models with up to 120 billion parameters entirely on-device, alongside a new Surface Laptop Ultra. The joint keynote signals that local inference is becoming the next PC battleground.
Meta and Microsoft are steering employees away from Anthropic's Claude coding tools toward in-house alternatives, The Information reported. Microsoft has cut its projected internal Claude spend by over a third from a planned $1 billion, while Meta's internal Claude Code users fell from about 60,000 to 30,000 as the company pushes its Muse Code and MetaCode tools. The shift reflects rising AI software costs and data-privacy considerations ahead of Anthropic's anticipated IPO.
Google DeepMind launched EmbeddingGemma 2, a 740-million-parameter open-weight embedding model that natively maps text, code, images, audio and video into one shared space for on-device search and retrieval. Released under Apache 2.0 and built on the Gemma 4 architecture, it leads sub-1B multimodal embedders on MTEB Code and MAEB benchmarks, runs in as little as 191MB of RAM for text-only use on a Pixel 11 Pro, and supports an 8K context window covering up to 5.5 minutes of audio or 58 video frames at once.
Embeddings are the quiet infrastructure of AI search: whoever owns the best on-device embeddings decides what private, offline retrieval looks like for a billion phones. A best-in-class open model at this layer pushes the whole ecosystem toward local-first AI, where your data never has to leave your device to be searchable.
Physicists at the University of Rome Tor Vergata built a physics-constrained diffusion model that synthesizes fully developed 3D turbulence, baking constraints such as incompressibility and prescribed mass and momentum fluxes directly into the generative dynamics. Tested on rotating turbulence, it stably reproduces anisotropic energy spectra and intermittent statistics, while standard denoising diffusion models show multiscale statistical deviations, break physical consistency and train substantially more slowly.
Turbulence has resisted both theory and simulation for a century; a generative model that respects exact physical laws rather than approximating them points the way for AI across physics, from climate to plasma. It also answers a growing worry that diffusion models are impressive interpolators with no respect for reality.
CaMeL protects a single agent from indirect prompt injection by separating trusted control flow from untrusted data, but the authors show these guarantees do not compose when agents invoke other agents as tools: they construct an attack that succeeds even when every agent runs CaMeL, because untrusted data gets reinterpreted as trusted input downstream. They propose multi-CaMeL, an agent-to-agent protocol that carries provenance across agent boundaries through a separate data channel, and introduce MultiAgentDojo to measure the security-utility trade-off.
Everyone is racing to deploy teams of agents, and the security story has been written for one agent at a time. A concrete proof that the best-known defense breaks under composition is the kind of result that changes how agent platforms get built, before the breaches force it.
A cross-modal contrastive multiple instance learning framework that imputes immunotherapy-associated molecular signatures directly from standard H&E slides, aligning visual morphology with molecular phenotypes in a shared latent space. Pathologists can query a whole slide image to surface transcriptomically coherent neighbors and approximate a costly 10-gene RNA signature for gastric adenocarcinoma without sequencing, supported by interpretable attention heatmaps.
It turns the most routine slide in pathology into a molecular pre-screening tool: no new assay, no extra tissue, just a smarter read of what is already on the glass. A retrieval-first design also fits how pathologists actually work, by case comparison rather than black-box scores.
MedPrune treats multi-agent medical VQA as a heterogeneous communication graph, where specialist agents are nodes and their interactions are edges, then dynamically prunes both with reinforcement learning: node sparsification drops task-irrelevant specialists, edge sparsification keeps only diagnostically salient links. The result is a medical multi-agent framework that reasons better while spending fewer tokens on redundant chatter.
Medical agent systems keep adding specialists until the token bill explodes; learning which experts and which conversations actually matter is the difference between a demo and something a hospital could afford to run.
A stylized reliability model for when multi-agent LLM systems actually help: decomposition saves attention cost by resetting context but pays a handoff tax when information is compressed between agents, while parallel sampling helps only insofar as failures are not shared. It yields two crossover conditions, tested on a ledger-reconciliation task, that reconcile the field's contradictory findings about multi-agent gains.
The literature cannot agree whether more agents help or hurt, because everyone is measuring a different bottleneck. A shared model of the trade-offs gives builders a way to predict, rather than argue about, when to split work across agents.
FairProp studies group fairness in graph neural networks, whose message passing can amplify topological bias. It bounds the demographic parity gap for node classification, link prediction and node regression across any number of sensitive groups, including the first bounds on deployed, sigmoid-activated link predictions and on node regression, and traces bias to two sources: separation of group means and within-group covariance of the representations.
GNNs decide who gets recommended, hired or flagged in networked data; fairness theory that covers the actual deployed prediction, not a proxy, is what makes debiasing auditable rather than aspirational.
AuraSE tackles hallucination in generative speech enhancement, where cleaner-sounding output can silently change words or speaker identity. A double-stream-to-single-stream multimodal diffusion transformer lets transcript and acoustic representations interact while keeping a dedicated path for the degraded input, and an inference-time policy optimization picks the decoding that best preserves content.
A hearing aid or call system that invents words is worse than a noisy one; making generative enhancement content-faithful is the step that lets these models leave the lab for ears that depend on them.
A post-training method for diffusion language models that minimizes the maximum mean discrepancy between generated and reference distributions in the feature space of a frozen pretrained diffusion LM. Token-level contextual features give multiple observations per sequence from a single pass, optimized with policy gradients for discrete models and direct differentiation for continuous ones.
Diffusion LMs promise faster, more controllable text generation, but aligning them has lagged behind autoregressive models; a distribution-matching objective computed in representation space gives them a practical post-training recipe.