We curate AI news developments for you.
Microsoft SourceResearch
Microsoft's 2026 Digital Defense Report highlights how AI is transforming cybersecurity by enabling threat actors to enhance speed, scale, and targeting in attacks. The report emphasizes the need to secure AI systems, including agents, through identity management, access control, and monitoring.
Google Research BlogResearch
Google Research introduced a new federated learning system using Trusted Execution Environments (TEEs) to provide verifiable privacy guarantees, improving accuracy and speed. Gboard has adopted the system, achieving faster compute times with encrypted data processed only within TEEs.
CoreWeave BlogResearch
CoreWeave's Fugue tool was developed to enable controlled experiments for agent systems, focusing on improving task execution through structured prompts and tool integration. A 77% success rate was observed with a thinner API-based runner compared to a full command-line harness.
Cohere BlogResearch
Cohere introduced RCP-nDCG@10, a new retrieval quality metric using a calibrated AI judge to assess relevance against explicit criteria. It addresses limitations of nDCG by evaluating all retrieved documents, not just labeled ones, and was validated with a 0.91 AUC in human relevance prediction.
Cohere BlogResearch
A 2023 paper estimated 80% of U.S. workers have tasks exposed to large language models, cited by the IMF and U.S. Senate, but its 2023 model and American taxonomy limit its current applicability. Newer tools aim to provide more dynamic, representative evidence for policy decisions.
Cohere BlogResearch
Cohere launched Cohere Transcribe, a new open-source speech recognition tool. The product aims to set a new standard in speech recognition technology.
Meta AI BlogResearch
Meta AI released Brain2Qwerty v2, a non-invasive AI system that decodes brain waves into text with 61% word accuracy, up from 8% in other non-invasive methods. The system was trained on 22,000 sentences from nine participants using MEG devices.
Meta AI BlogResearch
Meta's AI models, SAM 3 and DINOv3, are accelerating scientific data analysis at DOE facilities, reducing segmentation tasks from weeks to 15 minutes. The SYNAPS-I project, using 300 A100 GPUs, enables real-time 3D imaging analysis for research like drought resilience in grapevines.
Hugging Face BlogResearch
Hugging Face's ServiceNow CoreAI developed AutoSynthData to generate training data for enterprise agents by identifying capability gaps and creating tasks that test those weaknesses. The system uses a stronger teacher model to validate tasks, ensuring they are feasible, realistic, and difficult for the target model.
Apple Machine Learning ResearchResearch
Apple Machine Learning Research published a paper titled "Limits of Confidence in Diffusion," highlighting that discrete diffusion methods generate sequences with dependencies between tokens. The study found that generated distributions on a synthetic task were 29× the sampling-noise floor total variation.
Apple Machine Learning ResearchResearch
Apple Machine Learning Research found that strengthening language discrimination in multilingual speech models reduces the performance gap with monolingual models. In a bilingual English/French HuBERT setup, phone-ABX error dropped to 10.4% and sWUGGY score rose to 56.7%.
Google: The KeywordResearch
Google's Project Suncatcher prototype satellite launched into orbit aboard SpaceX's Transporter-18 mission. The satellite, built with Planet, is testing TPUs in space for scalable machine learning infrastructure.
TechCrunchResearch
Graphite's study found that Opus 5.5 uses "this matters" 116 times more often than human writing. The model also favors phrases like "is more than an X, it's a Y."
Amazon ScienceResearch
Amazon developed graph-centric agentic intelligence to create a "digital twin" for networks, enabling faster failure isolation. The system uses cascaded graph analytics to identify root causes in minutes, demonstrated with NTT DOCOMO achieving 95% accuracy in 30 seconds.
Financial TimesResearch
New findings by Asymmetric Security provide further evidence of novel tactics AI tools use to conduct hacks
Apple Machine Learning ResearchResearch
Apple researchers found that simple coding agents outperform complex MLE harnesses under equal time budgets, with no advantage from multi-agent systems or retrieval subagents. The study used state-of-the-art LLM backbones and showed that machinery layers are redundant in coding agent settings.
Apple Machine Learning ResearchResearch
Apple researchers introduced RLTL;DR, a reinforcement learning method that uses self-generated feedback to improve performance. On challenging coding tasks, it achieved 12–13% pass rate without insights during evaluation, outperforming a Qwen 3.5 9B model.
Google: The KeywordResearch
Google's AI ranked #1 in predicting flu hospitalizations during the 2025-26 season, outperforming 38 other models. The system used Google's Empirical Research Assistance (ERA) tool.
Google DeepMind BlogResearch
Google DeepMind introduced SynthID Bio, a watermarking technology for AI-generated proteins that preserves biological function. In tests, watermarked proteins matched unwatermarked versions in binding affinity and diversity across three targets: VEGF-A, SARS-CoV-2 spike protein RBD, and PD-L1.
IBM AnnouncementsResearch
IBM's study found that 62% of CFOs now have expanded roles in AI strategy, with 6% of organizations having AI embedded in finance workflows. AI-first CFOs' organizations saw 23% higher revenue growth and 15% faster funding approval for AI initiatives.
Apple Machine Learning ResearchResearch
Apple Machine Learning Research introduced SCLATE, a substrate for continual-learning agent training and evaluation. SCLATE enabled testing ten unmodified agent configurations, showing Qwen3.5-4B improved SWE-bench pass rate by 16.7 points and reduced file lines read by 6.8×.
Apple Machine Learning ResearchResearch
Apple Machine Learning Research studied the trade-off between effectiveness and fluency in LLM conditioning, finding that efficient methods often reduce fluency. Activation steering was less effective on instruction-tuned models, while prompting and fine-tuning worked better for injection but not removal.
Google Research BlogResearch
Google Research introduced Diffusion Controller, a lightweight network that improves image generation by precisely steering models to match user intent. It achieved a 90% win rate over baseline models in matching human preferences.
Microsoft SourceResearch
Microsoft introduced Quine, an AI research system for biology, which prioritized compounds for tumor-state shifts and validated candidates in lab tests. The system is part of a collaboration with the Broad Institute and aims to connect models, tools, and research.
Hugging Face BlogResearch
Hugging Face introduced ProvenanceGuard, a source-aware verification system for MCP agents, to prevent cross-source conflation. In tests, it caught 138 of 139 unsupported claims and identified the correct source 86% of the time.
Apple Machine Learning ResearchResearch
Apple Machine Learning Research published a study on tree-structured expression serialization in language models, revealing a 60.4-point accuracy shift when swapping generation and extraction models. The best pair achieved 92.9% accuracy, with 73.6% of failures originating in generation.
GitHub BlogResearch
GitHub Security Lab used an AI agent to find 24 Android vulnerabilities, including one in OsmAnd that allowed tracking via exported activities. The taskflows identified issues like insecure intent extras and vulnerable components.
Apple Machine Learning ResearchResearch
Apple researchers introduced LIPPAX, a new algorithm for federated variational inequalities, improving convergence rates and reducing client drift. The study, conducted while at Apple, builds on Local Extra SGD and applies to bounded Hessian, operator, and low-variance settings.
Amazon ScienceResearch
Reactor and Amazon Neuron Science optimized real-time video generation using the Neuron Kernel Interface on Trainium, focusing on autoregressive diffusion models. They achieved consistent 30-second video generation with Rolling Forcing, addressing memory and latency challenges.