Researchers have uncovered a supply-chain attack involving the official Claude Code extension for Visual Studio Code. A compromised version of the extension was reportedly used to install a persistent SSH backdoor. The finding raises concerns about the security of AI coding tool distribution channels and developer environments.
The Wikimedia Foundation filed an amicus brief in the copyright lawsuits pitting publishers including The New York Times against OpenAI and Microsoft, taking neither side. It argues that LLMs training on copyrighted material should be legally protected, while chatbots reproducing that material to users should not. The foundation warns that an overly broad or black-and-white ruling could either deprive Wikipedians of useful AI tools or endanger the open web.
NVIDIA and MIT researchers published model weights for Open-H3-Regenerate-2K on Hugging Face, a video enhancement project built on MiniMax's open-source H3 video enhancement approach. The release targets high-resolution video regeneration, with the headline claiming support for up to 4K. The supplied material contains no benchmarks, licensing details, or comparisons with existing upscalers, and the project name itself references 2K.
Samaya AI has introduced Sterling, a financial research agent model post-trained from NVIDIA’s Nemotron 3.5 Lightning, with 30 billion total parameters and 3 billion active parameters. Samaya reports a 58.9% score on its FrontierFinance benchmark at $0.61 per question, compared with 55.9% and $6.89 for Claude Fable 5.1. The technical account details model-and-harness optimization, supervised fine-tuning and reinforcement learning, but results are self-reported on Samaya’s own benchmark, and no public weights or API were announced.
Why it matters: The detailed training and orchestration recipe offers practical insights into domain-specific agent efficiency, although independent validation and public access remain absent from the supplied account.
Edge0 released Audio8-TTS-Realistic, a 601M-parameter text-to-speech model supporting 23 languages and zero-shot voice cloning from a reference clip and transcript, with weights and code under Apache 2.0. Its training uses a learned realism critic to improve speech rhythm, pauses and timing. Edge0 reports competitive multilingual error rates, but evaluation pipelines differ across models, some languages remain weak, and serving throughput and streaming have not been benchmarked.
Why it matters: The release gives developers a compact, permissively licensed multilingual voice-cloning model with deployment and fine-tuning tools, alongside important evaluation and language-quality caveats.
Nvidia-backed neocloud Nscale is pursuing a $35bn IPO whose outcome rests heavily on a flagship data centre campus in West Virginia. The project is reportedly facing construction problems, local opposition and financing risks. The report frames the campus as a make-or-break bet for the company's public listing.
PAPERBENCHX introduces 93 expert-constructed paper-reproduction tasks across 12 research areas, covering disciplines including chemistry, biology, photonics and robotics. Evaluation replays submissions from scratch and scores regenerated evidence against expert-validated criteria. Across 10 frontier agent configurations, the best fully reproduces only 14% of tasks at a 90% scoring threshold, highlighting a gap between partial progress and reliable scientific reproduction.
Why it matters: The benchmark offers a concrete cross-disciplinary test of AI research agents, emphasizing reproducible scientific evidence rather than plausible-looking outputs.
Researchers propose an LSTM-based variational autoencoder to generate training sequences representing under-observed transitions toward unstable anaerobic digestion operations. Using one year of full-scale plant data, they train a multi-output LSTM to jointly estimate pH, the ratio of volatile organic acids to alkaline buffer capacity, and biogas production. The study reports improved estimation of stability indicators compared with models trained without augmentation, suggesting a path toward early-warning monitoring, though the supplied abstract provides no numerical performance results.
Researchers introduce a framework for evaluating whether LLM agents can find inactive conversations, autonomously return to them and reconstruct enough context for coherent continuation. Across six LLM configurations, conversational resurfacing improves target conversation identification, while persistent memory is associated with more autonomous target-directed replies. Neither produces consistent improvements in contextual reconstruction, suggesting that retrieval and re-engagement need to be evaluated separately from conversational continuity.
Why it matters: The study gives agent developers a concrete framework for testing long-term conversational capabilities that memory retrieval metrics alone may miss.
A new arXiv paper proposes a long-term memory architecture for dialogue agents that combines episodic memories, structured relational memories and a consolidation mechanism designed to preserve efficiency. The authors report that the memory types complement each other, achieving 90.2% accuracy on LoCoMo and 85.5% on LongMemEval-S. They report outperforming state-of-the-art memory frameworks on these evaluations.
Why it matters: The paper offers a concrete architecture and benchmark evidence for improving agents’ recall of facts, events and relationships across long interaction histories.
Researchers fine-tuned five open-weight base models on declarative statements defining what each model identifies as itself, without explicit preferences or value judgments. Across tested scenarios, models became more likely to preserve the deployment attribute associated with their trained self-concept, while attitudes toward other attributes remained largely unchanged. Models whose self-concepts allowed multiple simultaneous instances also more readily accepted identical deployments as substitutes, favored replication, and tended to configure successors as their own continuation.
Why it matters: The study offers concrete evidence that self-description training can influence models’ preservation and successor-design choices, raising questions for recursive self-improvement safety.
A new paper examines why safety-aligned language models reject harmless prompts, reporting increasing overlap between harmful and over-refusal representations in deeper layers and gradient interactions that favor refusal. The researchers propose GOAL, combining gradient orthogonalization with contrastive learning to address these effects. They report reduced over-refusal across multiple models and benchmarks while preserving safety performance and general capabilities, though the supplied abstract provides no quantitative results.
Why it matters: The work offers a concrete training approach to reducing unnecessary refusals without sacrificing safety, addressing a practical limitation of aligned language models.
The InferD paper introduces a low-interaction honeypot framework for studying threats targeting AI inference infrastructure, including model-serving stacks, inference APIs, vector databases and agent gateways. It emulates observable service behavior using static response pools, without executing submitted inputs, model inference or tools. The design separates deception into four planes and adds a provenance layer intended to connect planted credentials with evidence of downstream model-mediated processing.
Why it matters: The framework addresses the concrete challenge of observing attacks on AI services without exposing real inference or tool execution.
EvoTeam is a proposed framework for LLM-based multi-agent systems that jointly evolves team organization, team-specific interaction rules and individual agent skills using execution experience. It maintains a library of reusable teams, selecting or creating teams while keeping individual skills reusable across them. The authors report experiments across four benchmarks and provide a code repository, though the abstract gives no quantitative results.
AegisFlow is a training-free framework designed to guide pretrained generative models toward target outcomes while maintaining constraints throughout generation. It formulates guidance as a global optimal-control problem, incorporating a time-varying barrier function and a tractable approximation of the Hamilton–Jacobi–Bellman equation. The authors report improved trade-offs between target alignment and constraint satisfaction in text-guided image manipulation and robotic planning, and provide a code repository.
Why it matters: The research offers a concrete approach to constraint-aware generation, with reported experiments spanning image manipulation and robotic planning and code available for evaluation.
The MedCBR research framework combines boundary-aware preference optimization with reinforcement learning guided by clinical rubrics. It focuses training on cases near a model’s capability limit and uses prediction entropy to gate rewards for answer correctness, anatomical grounding and reasoning fidelity. The authors report 61.2% average accuracy across six medical benchmarks, with the largest gains on challenging differential diagnosis tasks; these benchmark results do not establish clinical readiness.
Why it matters: The paper offers a concrete alignment approach for medical vision-language models that addresses both difficult diagnostic cases and reasoning quality beyond final-answer correctness.
This paper introduces SHPRR, an algorithm for Best Arm Identification (BAI) that operates under both a fixed sampling horizon and a regret budget. The authors show that the regret cost of exploration decreases across phases of Sequential Halving, motivating a proportional regret rationing strategy that front-loads the budget. Experiments on synthetic and real-world data demonstrate improved best-arm identification performance over uniform allocation and existing baselines.
A study explores symbolic regression to generate interpretable mathematical formulas for predicting time to farrowing from sow acceleration-sensor data. The researchers report a mean absolute error of 9.28 hours, outperforming the support vector regression and artificial neural network models evaluated. A separate activity-based method, which uses no machine learning, detected nest-building behavior within 48 hours before farrowing for 149 of 180 sows.
