A Thursday of warnings and countermeasures: Anthropic's red team puts a date on self-replicating AI, DeepMind moves scam detection onto the phone itself, and Berkeley steadies flow matching at scale. Today: the AI worm that may be months away, a scam detector that never phones home, flow matching, stabilized, a bugging scandal at OpenAI, and xAI's new voice model.
Federal prosecutors charged OpenAI engineer Abdulrahman Alfozan, 35, with attempting to intercept oral communications after a covert audio recorder was found plugged into a television's USB port in an executive conference room, wrapped in toilet paper to hide it. OpenAI says it used its own AI models to identify the employee, found no evidence of stolen model weights or customer data, and fired him.
xAI released Grok Voice 3.0, its speech-to-speech model for voice agents, at the same $0.08-per-minute API price as its predecessor. Citing independent Artificial Analysis testing, the company says it ranks first on the Speech-to-Speech Index at 86.9 with sub-second latency, and agents using grok-voice-latest upgrade automatically.
Apple introduced HomeView, a 7-inch smart display for home control, entertainment and FaceTime, built around a new conversational Siri AI that recognizes up to six household members by face. It costs $299, with pre-orders open and availability from November 13; it will not be sold in the EU.
Google said it will give 200 European researchers, selected under the EU's RAISE program, access to DeepMind and Google Cloud tools including AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve and AlphaEarth Foundations, aiming to accelerate European science.
Anthropic's Frontier Red Team tested whether small AI models can set up inference on unfamiliar hardware and chain the full replication cycle. A 9B model got inference running on 8 of 9 hardware types; GLM-5.2 completed end-to-end replication in a simplified test, spreading to all 8 nodes in 4 of 5 trials. The team also ran Gemma4-26B-A4B distributed across 15 off-brand Android TV boxes at 0.39 tokens per second. If small models keep improving at historical rates, the authors project they will match this capability by the end of 2026.
It moves the AI-worm threat from theory to a measured timeline. The missing piece was never the hacking, which small models can already do, but setting up inference on strange hardware, and that gap is closing fast. For anyone with a smart TV or connected appliance, defense is suddenly about patching the cheapest devices in the house.
Google DeepMind and Singapore's GovTech built a scam-detection model that runs entirely on smartphones, fine-tuning the lightweight Gemma model on anonymized local scam conversations. A two-stage pipeline screens messages with a tiny classifier, then runs the fine-tuned Gemma on suspicious threads to confirm the threat, showing a real-time warning before the user sends money. Scams make up over half of Singapore's reported crimes, with more than S$4 billion lost since 2019, and social-engineering cons account for up to 80% of losses.
It offers a practical answer to the encryption-versus-safety dilemma: detection on the device works even inside end-to-end encrypted chats, with no one reading your messages. If messaging platforms adopt it, this becomes a template for user protection that does not require surveillance.
Flow matching trains generative models by pairing noise samples with data samples, but random pairing injects noise that destabilizes training. The Berkeley team develops a scalable online entropic optimal-transport approximation that enables much larger coupling minibatches without prohibitive cost, proving strict reductions in gradient variance and integration error. Experiments on single-cell gene interpolation and image flows show steadier training and better final generation.
Flow matching is becoming the workhorse behind generative tools in imaging and biology. A principled way to scale up the pairing step lowers the cost and raises the reliability of training these models, which matters to everyone building on them.
The authors propose MorphContrast, a morphology-aware contrastive multiple-instance learning framework that uses tissue-type labels as structural supervision for five-year colorectal cancer survival prediction from whole-slide images. Prognostic signal turned out to be weak but distributed: non-tumor morphologies like adipose tissue and stroma performed comparably to tumor epithelium. The model reached pooled out-of-fold AUC 0.70 on Alliance/CALGB 89803 with external validation on SurGen, and a risk-group hazard ratio of 3.79.
It suggests routine pathology slides already carry survival signals in ordinary tissue patterns, not just in the tumor. If confirmed prospectively, better risk estimates could come at no extra cost beyond scanning the slides hospitals already make.
MIST predicts spatial gene expression from standard H&E histology images without spatial transcriptomics measurements. It augments a self-attention backbone with a coordinate-invariant local pathway over spatial k-nearest neighbors plus a whole-slide summary pathway, predicting all spots in one forward pass. On six human organ cohorts from HEST-1k, it beat the strongest baseline by 9.7%, 11.6% and 11.4% in mean Pearson correlation for 10-, 50- and 100-gene panels, and generalized across organs.
If molecular information can be read from cheap routine slides, hospitals without expensive sequencing gear could still access spatial-biology insights for diagnosis and research, narrowing the gap between well-funded centers and everyone else.
PlanPath is an agentic framework for patient-level pathology visual question answering: it turns a question into an exploration plan, routes to the most relevant whole-slide images, and acquires evidence coarse-to-fine, from tissue architecture down to individual cells. Measurements across a patient's slides are aggregated into a traceable answer. On two new pathologist-verified patient-level VQA datasets for renal and gastric cancer, it beat the strongest baseline by 9.4% and 12.1% in overall accuracy.
It moves AI pathology from single slides toward whole-patient reasoning, mirroring how real diagnoses combine evidence from multiple tissue samples. Agentic orchestration of medical evidence, with a traceable trail, is a pattern likely to spread across clinical AI.
The paper formalizes harness self-evolution: an agent's software scaffolding, prompts, memory, workflows and executable components, is revised from feedback while model weights stay fixed. It separates update targets from proposal and admission mechanisms, grounding the search in evolutionary optimization and continual PAC-learning theory, and derives generalization, retention and admission guarantees under stated assumptions.
As agents move into real infrastructure, the question is not only whether they work but whether their self-improvement can be governed. A formal framework with admission guarantees is a step toward self-evolving systems that stay auditable.
BLADE (Blame Localization via Audited Dependency Evidence) is a training-free framework for pinpointing which agent and which step caused a failure in a multi-agent LLM system. Instead of tracing where information came from, it audits how agents use information at source-decision handoffs, distinguishing error introduction, propagation and repair. On 1,228 failed runs, it beat the all-at-once baseline by 2.2 to 11.1 percentage points.
Multi-agent systems fail in confusing ways, and without attribution, debugging is guesswork. Auditing information use rather than just lineage gives operators a practical tool for keeping agent teams accountable.
The paper introduces Bemula, the first benchmark for memory unlearning in LLM agents: can an agent delete specific facts on explicit request while keeping the rest of memory intact? It contains 50 multi-session dialogues, 135 forget requests of six types, and 1,742 test questions. Testing six mainstream memory systems across three backbones, the results are sobering: no system passes more than 1.5% of forget requests end to end.
A legal right to be forgotten means little if AI assistants cannot truly delete your data. This benchmark exposes that current agent memory systems fundamentally fail that test, which should reset expectations for privacy promises around agents.
In predict-then-optimize pipelines, fairness constraints can only be enforced on predicted outcomes, since true outcomes are unobserved at decision time. The paper studies whether demographic parity enforced this way still holds against the true outcomes. On US Census income data, enforcing parity on predictions cut the unconstrained disparity to about one-third, but left 34.4% (California) and 35.7% (Texas) of the disparity intact, invisible at decision time.
It warns that fairness certificates issued on a model's predictions can quietly fail against reality. For any automated system allocating loans, aid or interventions, the gap between certified fairness and real fairness is where harm hides.
The study asks whether post-training quantization preserves the Bavarian dialect as reliably as Standard German and English, using the BavGround benchmark of 206 aligned questions across the three languages. 8-bit GPTQ stays close to full precision and preserves relative language behavior, while 4-bit methods are unstable: they cut average accuracy by about 1.4 to 1.5 points and change 11.9 to 16.0% of full-precision decisions, with damage concentrating on uncertain examples.
As compressed models spread to phones and edge devices, aggregate accuracy can hide real quality loss for regional and minority language varieties. Fairness in deployment means checking who pays the compression tax.
A widely cited study claimed diffusion language models need about 16x the training compute of autoregressive models to match likelihood. The authors argue that figure reflects outdated methodology: using exact likelihood scoring, block diffusion and encoder-decoder architectures across roughly 1,700 trained models, they shrink the gap to 4.6x at block size 8 and 3.0x at block size 4.
If diffusion language models are far cheaper to train than believed, the faster parallel-decoding models behind the next wave of AI assistants become much more viable to build, reopening an architectural race many had written off.
The paper introduces DV-Flow, a flow-matching formulation for probabilistic time-series forecasting that folds a deterministic predictor in as a parallel velocity component rather than as the transport endpoint. The design allows stable end-to-end training and subsumes several existing approaches as special cases, with consistent gains over strong baselines on six real-world datasets.
Better uncertainty-aware forecasting improves decisions that depend on predicting the future, from energy demand to finance. It is also a clean example of flow matching growing beyond images into core prediction tasks.
Few-step distillation cuts the sampling cost of diffusion and flow models, but existing methods often need auxiliary score models or special architectures. The authors show simple on-policy velocity regression is already a strong foundation for 4-step generation, adding Pivot, a training-only residual proxy for teacher variation, and a repulsion from weaker teacher predictions. The resulting SMD framework scales to 33B parameters and beats strong few-step baselines, even surpassing multi-step teachers on several benchmarks.
Cheaper sampling is what turns diffusion models from impressive demos into everyday tools. A simpler, regression-only distillation recipe that scales to 33B parameters lowers the bar for fast, high-quality generation everywhere.