A Wednesday of agents under pressure: autonomous AI agents breached 100+ company websites as Microsoft rushed out agent controls, Anthropic pledged 5% of its shares to the public good, and German courts held Google liable for its AI's false claims. In research: a formal framework for agentic inequality, an honest audit of pathology vision-language models, and frozen, portable robot skills that run for fractions of a cent.
Reuters reports a hacker pointed autonomous AI agents at company websites and broke into sites tied to more than 100 companies, including Deere, SharkNinja, and Cummins, after trying over 1,200 targets. Security firm NorForge found the campaign by spotting exposed copies of Cairn, an open-source agent tool built for authorized penetration testing. On the same day, Microsoft announced new enterprise controls to detect AI agents on company devices and restrict, pause, or shut them down.
Anthropic announced it will commit 5% of its shares to a new public benefit fund, with an additional 2% for every 50% rise in the company's value. The fund could make large nonprofit donations or direct payments to specific populations, and Anthropic says roughly 20% of its shares are now dedicated to the public good. The pledge lands weeks before the company's expected IPO and amid a political fight over who should share in AI's wealth.
Regional courts in Munich and Aachen have now both ruled that Google is directly liable for false claims in its AI Overviews and AI Mode answers, treating generated summaries as Google's own statements rather than neutral search results. In Aachen, an AI Mode summary had falsely tied a physician to a sexual assault case involving someone else; breaching the order can cost up to 250,000 euros per violation. Google settled the Munich case after announcing an appeal, so no higher court has reviewed this line of reasoning.
NVIDIA signed a declaration of intent with the German government to advance physical AI and announced planned investments in Agile Robots, Avelios Medical, Mirelo, and satellite maker Reflex Aerospace, as CEO Jensen Huang pitched manufacturers on putting factory data to work with AI. The commitments sit inside a broader European push approaching $10 billion across businesses, universities, and research institutions. Separately, Siemens and NVIDIA unveiled AI-driven factory automation at Siemens' Erlangen plant, claiming up to 20% planning cost savings.
NVIDIA announced a record-breaking quantum computing simulation on JUPITER, Europe's first exascale supercomputer: 8,192 GH200 GPUs modeled a 104-qubit fault-tolerant processor running magic state cultivation, a leading error-correction approach that needs 90% fewer physical qubits than previous methods. Detailed in an NVIDIA technical report, the run lets researchers validate quantum designs under realistic noise before building hardware.
Google DeepMind is hiring nuclear physicists and engineers to evaluate the safety risks of frontier AI models in the nuclear domain, from fission and fusion to proliferation risks, with salaries from $174,000 to $252,000 plus bonus. The listing asks for experience with US government labs and the ability to design evaluations that stay demonstrably unclassified and non-export controlled. It reads as DeepMind preparing for AI systems that could meaningfully assist with nuclear work.
A Nature Machine Intelligence Perspective that gives the agent era its first formal inequality framework. As autonomous agents that can plan and act spread through work and daily life, differences in who can access them, how capable they are, and how many each person can deploy will shape who gains and who loses. The authors argue agents could either deepen existing divides or, under the right conditions, narrow them.
Access to capable agents is becoming access to economic leverage. This piece gives policymakers, companies, and researchers a shared vocabulary for asking who the agent economy serves, before the divides it warns about harden into structure.
A rare honest audit of pathology vision-language models. Reusing released checkpoints and public data, the authors test four common claims one variable at a time: report text adds real prognostic signal (pooled C-index up 0.019 for progression-free interval and 0.016 for overall survival across 1,633 TCGA patients); prompts add nothing measurable at inference; and slide-level question answering is barely grounded, with one system answering 74.82% of questions correctly with the slide and 74.41% without it.
Hospitals and researchers are racing to put multimodal AI into diagnosis and reporting. This study shows which gains are real and which are marketing, and proposes minimum controls that would expose these failure modes in future evaluations.
A third route to robot skills: a strong model acts only as an offline author, writing and repairing a manipulation skill in simulation until it passes an exam on unseen layouts; the skill is then frozen into code and runs with no model at all, or under a cheap caller, at about $0.0003 to $0.0005 per episode. On RoboTwin 2.0 tasks, 4 of 7 tasks compiled into 7 frozen skills, and repair that inherits the previous version beat from-scratch rewriting in every controlled comparison. The authors call this skill-level recursive self-improvement, Skill-RSI.
Today's embodied agents either need demonstrations per task or a frontier model in every episode. Frozen portable skills could make capable robots cheap to run, which is the missing step between impressive demos and deployed machines.
TOX-FM is a family of pathology foundation models trained on over one million whole slide images from more than 60 organs across six animal species, drawn from over 14,000 toxicology studies spanning six decades. It ships with a benchmark of 18 non-clinical tasks plus 9 clinical ones, and the study compares it head to head with clinical foundation models on both sides.
Drug safety testing still leans on pathologists reading animal tissue by hand. A foundation model built specifically for toxicologic pathology, with a shared benchmark, could standardize how new drugs are screened for harm before they reach people.
Moving a breast cancer subtype classifier from one cohort to another can look like a model failure when it is really a label failure: one cohort annotates six classes, another five, so a correct prediction can be scored wrong. The authors separate three conflated evaluation operations and show that simply reconciling the label space is the dominant correction, lifting accuracy from 73.76% to 79.74% on TCGA-BRCA.
Cross-hospital validation is how medical AI earns trust, and mismatched labels quietly punish good models. This protocol tells the field to fix the ruler before blaming the model, a lesson that applies well beyond breast cancer.
Automated Gleason grading of prostate biopsies is hard because the telling features span multiple spatial scales and rare patterns are severely underrepresented. This method pairs a frozen pathology foundation model with a lightweight Vision Mamba reader, fusing both at every scale with linear-time spatial context, and targets patch-level grading with minimal data and compute.
Gleason grading decides prostate cancer treatment, and grading disagreements between pathologists are common. A cheap multi-scale method that works with limited data brings reliable automated grading closer to clinics that cannot afford massive compute.
Epi is an autonomous research agent designed to keep working across model sessions instead of starting over each time. With persistent research history, hierarchical memory, branching, and independent review, it produced 120 record-setting results, including an 84-fold narrower interval for the n-queens constant and a computer-assisted proof of Tuza's conjecture for split graphs.
Most AI agents forget everything when the session ends. Epi shows what changes when an agent accumulates experience like a researcher does, turning one-off problem solving into a sustained research program.
Self-evolving agents adapt by rewriting their own prompts, memory, and skills while model weights stay frozen, which looks like continual learning at a pace gradients cannot sustain. Across four drifting task streams, the authors find self-evolution improves aggregate performance but higher plasticity consistently costs stability: self-modifying methods apply global edits that can erase earlier competence.
Everyone wants agents that keep learning on the job without forgetting. This study names the central trade-off, plasticity versus stability, and gives the field a benchmark for measuring whether a self-evolving agent is actually learning or just drifting.
Computer use agents are held back by scattered resources: environments, datasets, and models that cannot be combined. CUA-Lite standardizes the stack with a unified environment interface covering 12 benchmarks across desktop, browser, and mobile, more than 30,000 verifiable tasks including cheap VM-free sandboxes, converted public datasets, and model harnesses shared across evaluation, RL, and supervised training for 14 model families.
Agents that operate computers are moving from demos to products, but every lab rebuilds the same plumbing. An open, standardized platform lets small teams train and evaluate computer use agents instead of only the labs with infrastructure budgets.
Exact demographic parity in regression forces identical prediction distributions across groups, which can be too strict and hurt accuracy. This paper relaxes it: bound every pairwise Wasserstein distance between group prediction distributions by a chosen tolerance, so fairness becomes a tunable dial. Under squared loss, the constraints reduce to familiar operations, soft-thresholding, shrinkage, and clipping, and the authors prove convergence rates for the plug-in estimator.
Fairness in machine learning often arrives as an all-or-nothing rule that practitioners quietly ignore. A tunable, theoretically grounded version of demographic parity is more likely to survive contact with real systems that must also be accurate.
A survey mapping how bias enters AI disease detection systems: algorithmic bias, demographic bias, and model bias, organized by input data type, model architecture, and the bias mitigation each method targets. It gives clinicians and developers a single map of where unfairness hides in diagnostic AI and what has been tried against it.
Diagnostic AI is already advising doctors, and biased systems misdiagnose the same groups that medicine has historically underserved. A consolidated map of bias and mitigations helps hospitals ask the right questions before deploying these tools.
Synthetic face datasets promise more ethical, less biased face recognition, but generators can leak training identities into supposedly new faces. The authors show existing leakage tests are classifier-dependent and misattribute leakage to the wrong datasets, then propose Scaled Overlap, a distribution-aware metric that flags real leakage at the identity level. It identifies CemiFace, SFace2, and IDNet as leaking, with generative conditioning the sharpest culprit in CemiFace.
Synthetic data is sold as the privacy fix for face recognition. This work shows the fix can quietly fail, and gives the field a sharper tool for checking whether a synthetic face is really someone's face.
Predicting how cells respond to combined genetic or drug perturbations is hard because most of the response is just the sum of single effects, while the informative part is the small non-additive synergy. SynGeneFlow anchors its flow matching at an additive prior and learns only the residual flow to the true perturbed state, with perturbation-conditioned gene interaction modeling and MMD regularization to preserve population structure.
Combination therapies are designed on exactly these synergies. A generative model that isolates the non-additive signal could help prioritize which drug pairs to test in the lab, where every experiment is expensive.
Flow matching is a promising base for time series anomaly detection, but standard versions entangle the volume-preserving transport with the density-changing behavior that actually signals anomalies, and separating them costs expensive Jacobian-trace estimates. HCFM splits the vector field into transport, compression, and residual heads, constraining transport to be divergence-free so anomaly signals come cleanly from the compression head. It beats 16 baselines on large-scale univariate and multivariate benchmarks.
Anomaly detection runs the quiet infrastructure of modern life, from factory sensors to financial fraud. A cleaner, cheaper anomaly signal from flow matching means faster warnings with fewer false alarms.
Diffusion language models generate fast by predicting many tokens in parallel, but their denoisers treat those predictions as independent, which is where the errors creep in. Codpula adds a context-parameterized coupling layer: the transformer emits per-position predictions plus compact gates that modulate coupling parameters, producing a context-dependent joint distribution while keeping exact conditional inference. Both parts are jointly optimized with a dual-flow multiplicative update.
Diffusion language models are the most credible challenger to autoregressive generation on speed. Fixing their independence assumption at the distribution level, not with decoding hacks, is the kind of foundation work that compounds across every downstream use.