# backends Published articles for backends. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility DevFeed: [The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility]() Original publisher: [Read original article]() Author: David Pape, Jonathan Evertz, Lea Sch\\"onherr Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [reproducibility](), [Large Language Model](), [vllm](), [llama.cpp](), [LLM optimization](), [CUDA Graphs]() Tags: [abstract](), [adoption](), [alter](), [analyze](), [arithmetic](), [arxiv](), [backend](), [backends](), [benchmark](), [benchmarks](), [caching](), [can](), [choice](), [comparisons](), [cost](), [critical](), [cs-lg](), [impact](), [inference](), [llm](), [reproducibility]() ### AI overview Inference backend choice can materially affect LLM benchmark results even when model weights, decoding parameters, and hardware are held constant. Across five engines and open-weight models from 4B to 30B parameters, the study reports score shifts of up to 16.6 percentage points and frequent output disagreement. It attributes divergence to backend-specific optimizations and defaults, including prefix caching, CUDA graphs, custom kernels, and logit processing. A survey found 200 distinct engines, while analysis of 35,000 ML publications found that inference stacks are rarely reported. The authors argue that standardized reporting of inference stacks would improve benchmark reproducibility and interpretability. ### Source excerpt arXiv:2605.19537v3 Announce Type: replace Abstract: Progress in LLMs is increasingly measured through standardized benchmarks, where state-of-the-art improvements are often separated by fractions of a percentage point. At the same time, the computational cost of evaluating modern LLMs has driven widespread adoption of specialized inference backends, software systems that execute trained models efficiently at inference time. While critical for scalability, system-level optimizations, such as custom CUDA kernels and reduced-precision arithmetic, can alter token probabilities and introduce non-determinism, possibly cascading into divergent generation. In this work, we first survey the inference landscape, identifying $200$ distinct engines, and analyze $35,000$ ML publications, finding that the specific inference stack is rarely reported despite this widespread diversity. We then present a systematic empirical study of how inference backends affect LLM benchmark results. Holding model weights, decoding parameters, and hardware constant, we evaluate five widely used inference engines, including vLLM, SGLang, and llama.cpp, across multiple open-weight models spanning 4B-30B parameters and established benchmarks. We show that the choice of backend alone can shift benchmark scores by up to $16.6$ percentage points and induce high rates of output disagreement. By isolating backend optimizations and tracing the execution pipeline, we find this divergence is driven by system-level optimizations such as prefix caching and CUDA graphs, custom kernels, and engine-specific defaults in logit processing. Our findings identify the inference backend as a previously unreported yet consequential hyperparameter in LLM evaluations and advocate standardized reporting of inference stacks to improve the reproducibility and interpretability of benchmark comparisons. ## Java and the Next Generation of Programming Models DevFeed: [Java and the Next Generation of Programming Models]() Original publisher: [Read original article]() Author: Ana-Maria Mihalceanu Published: 2026-11-08T00:00:00Z Content type: article Language: en Sources: [Inside Java]() Topics: [Java](), [AI Chat](), [Programming Languages / Compilers]() Tags: [ai](), [ai-models](), [analyze](), [api](), [artificial-intelligence](), [backends](), [bridge](), [can](), [code](), [engines](), [express](), [external](), [files](), [function](), [gap](), [hardware](), [inference](), [java](), [jvm](), [models](), [programming](), [project-babylon](), [project-panama]() ### AI overview The session presents two ways to work with specialized runtimes and AI models from Java. For current inference work, it covers the Foreign Function and Memory API (FFM) and tools such as jextract to interact with runtimes outside the JVM. It then introduces Project Babylon's Code Reflection as a proposed path for expressing AI models as Java programs that can be analyzed, transformed, and lowered to inference engines and hardware accelerator backends, without external model files or glue code. The description frames these as present-day FFM-based inference and a future Code Reflection approach; it does not provide implementation details or performance results. ### Source excerpt Modern software increasingly interacts with AI models, query languages, and other specialized programming models, forcing you to switch contexts, embed foreign code, or write verbose wrappers to bridge the gap. What if you could express logic from outside the JVM directly in Java? In this session, we start by looking at how Java enables you to safely interact with runtimes outside of the JVM using the Foreign Function and Memory API and tools like jextract. Then will push further with Project Babylon's code reflection: express AI models as Java programs that Babylon can analyze, transform and lower to inference engines, thus leveraging hardware accelerator backends while skipping external model files or the need to glue code. ## Screen Before You Serve: Simulation for Production Customer Experience AI Agents at 140M Scale DevFeed: [Screen Before You Serve: Simulation for Production Customer Experience AI Agents at 140M Scale]() Original publisher: [Read original article]() Author: Edesio Alcoba\\c{c}a, Kevin Rossell, Aman Gupta, Shao Tang, Jiwoo Hong, Pabel Carrillo-Mendoza, Wanderson Concei\\c{c}\\~ao Ferreira, Alvaro Tedeschi, Zayd Simjee, Shreya Rajpal, Bruno Finardi Hime, Chri Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [AI for customer service automation](), [agentic workflows](), [Test automation](), [AI Chat]() Tags: [abstract](), [agent](), [agentic](), [agentic-workflows](), [agents](), [ai](), [ai-agents](), [also](), [arxiv](), [backends](), [brazil](), [can](), [cs-ai](), [cs-cl](), [customer](), [customer-experience](), [experience](), [production](), [scale](), [screen](), [serve](), [simulation]() ### AI overview The Snowglobe workflow screens customer experience agents before deployment by having synthetic customers react to agent responses and simulating tool outputs, enabling multi-step tests without production backends. On Nubank's Card Delivery and expanded Card Management agents, simulated and production binary evaluator scores were highly correlated across four deployed versions. Simulation-guided iteration raised transactional net promoter score by 36.69 points in a live A/B test. Screening open-weight configurations across more than 16,000 simulated conversations preceded a live test in which the selected model raised self-service rate by 8.82 percentage points to Nubank's highest observed level, with no statistically significant tNPS change. The abstract reports results for this Nubank deployment; it does not establish that the workflow generalizes to other agents. ### Source excerpt arXiv:2609.30137v2 Announce Type: replace Abstract: Customer experience (CX) agents use tools and large language models to address customer requests and guide conversational interactions with an organization's products. Improving these agents, especially in regulated industries, is difficult: they must detect intent, follow complex operational policies and use tools reliably. Manual end-to-end testing offers limited coverage, while live experiments expose customers to failures that can erode trust. We present a hypothesis-driven simulation workflow for screening candidate CX agents before deployment. Synthetic customers react to agent responses and simulated tool outputs enable multi-step agentic workflows without invoking production backends. We use the Snowglobe simulator on Nubank's Card Delivery agent and its expanded successor, Card Management - Nubank's highest-volume chat-support agent in Brazil. Across 4 deployed versions, simulated and production version-level binary evaluator scores show high correlation. Simulation-guided iteration increased transactional net promoter score (tNPS) by 36.69 points in a live A/B test. We also screened open-weight configurations in over 16,000 simulated conversations. In a subsequent live A/B test, the selected model increased self-service rate (SSR) by 8.82 percentage points to the highest level observed at Nubank, with no statistically significant change in tNPS. Simulation made broad exploration of models, reasoning settings, and prompts feasible without customer exposure, enabling production improvements that would have been impractical to pursue through live experimentation alone. ## Wave-Based Dispatch for Circuit Cutting in Hybrid HPC-Quantum Systems DevFeed: [Wave-Based Dispatch for Circuit Cutting in Hybrid HPC-Quantum Systems]() Original publisher: [Read original article]() Author: Ricard S. Raigada-Garc\\'ia, Josep Jorba, Sergio Iserte Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Distributed Computing]() Topics: [High-Performance Computing]() Tags: [backend](), [backends](), [circuit](), [circuits](), [cloud](), [code](), [concurrency](), [cpu](), [cs-dc](), [dispatch](), [execution](), [failover](), [fragment](), [fragments](), [framework](), [hpc](), [hybrid](), [quantum](), [quantum-systems](), [recovery](), [systems]() ### AI overview DQR makes circuit fragments first-class schedulable units so HPC centers can manage hybrid HPC-quantum workloads with existing resource policies. It uses backend-agnostic fragment descriptors and a wave-based coordinator with non-blocking polling, and was implemented on CESGA's Qmio system with local QPU and remote IBM Torino backends. In tests on a 32-qubit Hardware-Efficient Ansatz circuit, it improved makespan over a monolithic CPU baseline and rerouted fragments from the local QPU to classical simulators after hardware incompatibilities without restarting the pipeline. For deeper circuits, coordination residual was 5% of total execution time. These reported results suggest low coordination overhead in that evaluation; the supplied abstract does not give broader benchmark details. ### Source excerpt arXiv:2604.15279v2 Announce Type: replace Abstract: Hybrid High-performance Computing (HPC)-quantum workloads based on circuit cutting decompose large quantum circuits into independent fragments, but existing frameworks tightly couple cutting logic to execution orchestration, preventing HPC centers from applying mature resource management policies to Noisy Intermediate-Scale Quantum (NISQ) workloads. We present DQR (Dynamic Queue Router), a runtime framework that bridges this gap by treating circuit fragments as first-class schedulable units. The framework introduces a backend-agnostic fragment descriptor to expose structural properties without requiring execution layers to parse quantum code, a wave-based coordinator that achieves pipeline concurrency via non-blocking polling, and a production-ready implementation on the CESGA Qmio supercomputer integrating both QPUs local on-premises (Qmio) and remote cloud (IBM Torino) backends. Experiments on a 32-qubit Hardware-Efficient Ansatz (HEA) circuit demonstrate not only makespan improvements over a monolithic CPU baseline but also transparent per-fragment failover recovery-specifically rerouting tasks from the local QPU to classical simulators upon encountering hardware-level incompatibilities-without pipeline restart. For deeper circuits, the coordination residual accounts for only 5% of the total execution time, highlighting the framework's scalability. These results show that DQR enables HPC centers to integrate NISQ workloads into existing production infrastructure while preserving the flexibility to adopt improved cutting algorithms or heterogeneous backend technologies. ## A Closer Look at Phonon DevFeed: [A Closer Look at Phonon]() Original publisher: [Read original article]() Author: Gradium; Hugging Face Published: 2026-11-05T00:00:00Z Content type: article Language: en Sources: [Gradium]() Topics: [Open Source](), [hugging face](), [spaces](), [huggingface\_hub](), [Tool](), [kernels](), [open-llm-leaderboard]() Tags: [ai](), [announcement](), [apache-2-0](), [apache-2-0-tts](), [apple](), [apple-watch](), [apps](), [arabic](), [arm](), [audio](), [autonomous](), [backends](), [benchmark](), [browser](), [chat-template](), [checkpoint](), [cloud](), [codec](), [hugging-face](), [multilingual-tts](), [on-device-tts](), [open-tts-leaderboard](), [open-weight-tts](), [phonon](), [phonon-31m](), [phonon-40m](), [pocket-tts](), [ptts](), [time-to-first-audio](), [transformers-tts](), [tts]() ### AI overview Phonon is an open-weights text-to-speech model designed to run locally without a server or network, keeping typed text on the device. The release includes models of about 65 MB and 75 MB, more than 100 built-in voices, and support across English, French, German, Spanish, and Portuguese. Phonon 31m is Apache 2.0, including commercial use; Phonon 40m is non-commercial. The authors attribute the models' compact quality to flow matching and report that Phonon ranks first in German and Spanish intelligibility and in the top three in each evaluated language, while leading CPU streaming time-to-first-audio benchmarks. Their ptts framework targets single-stream inference across CPU, GPU, and NPU backends; the Transformers implementation targets GPU throughput with batched generation and full-graph torch.compile. Custom voices can be cloned from a short recording with the speaker's consent or designed in text. The leaderboard is described as new, and the reported rankings apply to the languages evaluated. ### Source excerpt Two weeks after opening up Phonon, the story of the launch: why we built it, what it can do, the demos that came with it, and how it did on Hugging Face's Open TTS Leaderboard. ## A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks. DevFeed: [A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.]() Original publisher: [Read original article]() Author: Olesia Krizhevich; Laurent Mazare; Eric Bezzam; Eustache Le Bihan Published: 2026-11-05T00:00:00Z Content type: article Language: en Sources: [Hugging Face]() Topics: [On-device AI](), [Open Source](), [Artificial Intelligence]() Tags: [ai](), [apache](), [apache-2-0](), [apple](), [apple-watch](), [apps](), [arabic](), [arm](), [audio](), [autonomous](), [backends](), [benchmark](), [browser](), [device](), [leaderboard](), [on-device](), [open](), [phonon](), [quality](), [speech](), [tts]() ### AI overview Phonon is a small text-to-speech model designed to generate speech locally without a server or network connection. Its 31m English model uses Apache 2.0; the 40m model supports English, French, German, Spanish, and Portuguese under a non-commercial license. The models are about 65-75 MB, include over 100 built-in voices, and support custom voices through cloning or voice design. The article reports first place in German and Spanish intelligibility and top-three placement in every evaluated language, alongside models up to 100 times larger; Phonon also leads the CPU streaming benchmark for time-to-first audio. Gradium's ptts framework targets single-stream inference across more than 10 CPU, GPU, and NPU backends. The Transformers PyTorch implementation instead targets GPU throughput through batched generation and torch.compile. The article also describes demos and integrations, and notes that leaderboard measures include intelligibility, speed, and naturalness. ### Source excerpt We're on a journey to advance and democratize artificial intelligence through open source and open science. ## ModelBlaster: Closing the Performance Gap through Learning-Augmented Agents Across Models, Kernels and Hardware Backends DevFeed: [ModelBlaster: Closing the Performance Gap through Learning-Augmented Agents Across Models, Kernels and Hardware Backends]() Original publisher: [Read original article]() Author: Shengjun Kris Dong, Dima Nikiforov, Sahil Modi, Sana Damani, Edward C Lin, Siva Kumar Sastry Hari, Christos Kozyrakis Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Reinforcement learning](), [GPU optimization](), [High-Performance Computing](), [OpenCL]() Tags: [2](), [3](), [abstract](), [accelerators](), [actions](), [actor](), [adreno](), [agents](), [algorithms](), [apache](), [architectures](), [arxiv](), [backends](), [base](), [cs-lg](), [gap](), [hardware](), [kernels](), [learning](), [models](), [performance]() ### AI overview ModelBlaster reuses optimization experience across kernels, whole models, and heterogeneous accelerators. Its persistent knowledge base maps performance states to actions, outcomes, and constraints, and is updated at inference time using Actor-Critic Textual Gradient Descent; a model-level optimizer draws on kernel-level knowledge. One optimization loop and knowledge schema support CUDA on NVIDIA GPUs, OpenCL on Qualcomm GPUs, and RVV intrinsics on RISC-V vector cores, while lowering and measurement remain target-specific. The abstract reports B200 geometric-mean speedups over PyTorch of 1.85x, 5.58x, and 5.18x on KernelBench Levels 1-3, and 1.69x on SOL-ExecBench Levels 1-2. It also reports 5.8x over the Qualcomm vendor SDK and 1.7x over Apache TVM on Adreno GPUs, plus 2.7x over ExecuTorch with XNNPACK across 70 KernelBench problems on RISC-V vector cores. ### Source excerpt arXiv:2611.04104v1 Announce Type: new Abstract: Optimizing accelerator code across hardware architectures is difficult. ML compilers search predefined spaces of schedules and micro-kernels, while LLM-guided optimization can explore alternative algorithms, operator decompositions, and hardware-specific techniques. However, existing LLM agents typically target a single programming interface, toolchain, or accelerator family, focus on kernel-level optimization without model-level transforms, and offer limited reuse of experience across hardware targets. We present ModelBlaster, a Memory-Augmented In-Context Reinforcement Learning (MAIC-RL) framework for optimizing data-parallel kernels and whole models across heterogeneous SIMD and SIMT accelerators. Its mutable policy is a persistent knowledge base updated at inference time by Actor-Critic Textual Gradient Descent (AC-TGD). The knowledge base maps performance states to actions, outcomes, and constraints, enabling experience reuse across kernels, models, and hardware targets. A model-level optimizer uses kernel-level knowledge to inform whole-model decisions. The same optimization loop and knowledge schema support CUDA on NVIDIA GPUs, OpenCL on Qualcomm GPUs, and RVV intrinsics on RISC-V vector cores, with only lowering and measurement layers specific to each target. Compared with PyTorch, ModelBlaster achieves geometric-mean speedups of 1.85x, 5.58x, and 5.18x on KernelBench Levels 1, 2, and 3, and 1.69x on SOL-ExecBench Levels 1 and 2 on a B200 GPU. On Qualcomm Adreno GPUs, it achieves 5.8x over the vendor SDK and 1.7x over Apache TVM; on RISC-V vector cores, it achieves 2.7x over ExecuTorch with XNNPACK across 70 KernelBench problems. ## Never lose robot data again: automatic ROS 2 bag recording and upload DevFeed: [Never lose robot data again: automatic ROS 2 bag recording and upload]() Original publisher: [Read original article]() Author: gbeuzeboc Published: 2026-11-03T16:02:04Z Content type: tutorial Language: en Sources: [Ubuntu blog]() Topics: [recording](), [snap]() Tags: [automatic](), [automatically](), [aws](), [aws-s3](), [backends](), [bags](), [central](), [ceph](), [cloud](), [command](), [configuration](), [content](), [fleet](), [recording](), [robot](), [robots](), [ros](), [ros-2](), [storage](), [upload]() ### AI overview The `ros2-exporter-agent` snap automates ROS 2 bag capture and upload: it records topics to MCAP files, splits bags by time or size, tries to sync them every two minutes through rclone, restarts recording daily, and deletes local files older than 10 minutes. Recording starts only after auto-clean is running, and sync retries if the server is briefly unreachable, leaving bags on the robot only as a short staging buffer. Uploads use an rclone remote named `bagstore`, which can alias an SFTP server or S3-compatible storage; the article says rclone supports more than 40 backends. Setup requires `rclone.conf` and `rosbag2-recorder.yaml` for storage and recording choices. For fleets, a configuration snap can distribute shared settings and updates, with an example that generates a device UID. Suggested extensions include copying bags with rclone, browsing them via a file browser, or opening remote MCAP files in Foxglove. ### Source excerpt If you work with robots, perhaps you've come up against a common, but annoying situation: the one field run that mattered is the one nobody recorded. Or it was recorded but the ROS 2 bag (a recording of ROS 2 topics) is sitting in the robot's storage, at a customer site, three time zones away. [...] ## Never lose robot data again: automatic ROS 2 bag recording and upload DevFeed: [Never lose robot data again: automatic ROS 2 bag recording and upload]() Original publisher: [Read original article]() Author: gbeuzeboc Published: 2026-11-03T16:02:04Z Content type: tutorial Language: en Sources: [Canonical]() Topics: [recording](), [snap]() Tags: [automatic](), [automatically](), [aws](), [backends](), [fleet](), [recording](), [robot](), [ros-2](), [storage](), [upload]() ### AI overview The `ros2-exporter-agent` snap automates ROS 2 bag capture and upload through four services: a recorder writes MCAP bags split by time or size, synchronization attempts uploads every two minutes with rclone and retries when the server is unreachable, daily rotation restarts recording at midnight, and auto-clean removes local files older than 10 minutes. Recording starts only after auto-clean is running, leaving bags as a short local staging buffer. Uploads use an rclone remote named `bagstore`, which can alias an SFTP server or S3-compatible storage; the excerpt says rclone supports more than 40 backends. Robots can be configured with local `rclone.conf` and `rosbag2-recorder.yaml` files, or receive shared configuration through a separate configuration snap for fleet-wide updates. The excerpt also suggests pulling bags with rclone, browsing stored files, or opening remote MCAP URLs in Foxglove. ### Source excerpt If you work with robots, perhaps you've come up against a common, but annoying situation: the one field run that mattered is the one nobody recorded. Or it was recorded but the ROS 2 bag (a recording of ROS 2 topics) is sitting in the robot's storage, at a customer site, three time zones away. [...] ## \\textsc{ChatDify}: Automating End-to-End Dify Workflow Synthesis from Natural Language via Multi-Agent Collaboration DevFeed: [\\textsc{ChatDify}: Automating End-to-End Dify Workflow Synthesis from Natural Language via Multi-Agent Collaboration]() Original publisher: [Read original article]() Author: Tao Fan, Guoqiang Ma Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Multi Agent Systems](), [AI Agent](), [Multi-step AI processes](), [Large Language Model](), [Orchestration](), [qwen]() Tags: [actor](), [agent](), [agent-framework](), [agents](), [application](), [application-development](), [applications](), [architecture](), [backends](), [benchmark](), [collaboration](), [compilation](), [context](), [cs-ai](), [deepseek](), [development](), [dsl](), [multi-agent](), [natural-language](), [synthesis](), [workflow]() ### AI overview ChatDify automates Dify workflow creation from natural-language requests through three collaborating agents: a Planner turns requests into execution plans, an Executor synthesizes parameters and compiles a schema-augmented DSL, and a Reviewer verifies results and refines them in a feedback loop. On 100 production-inspired tasks, reported success rates varied by LLM backend: 0.96 with Qwen-3.7, 0.91 with GLM-5.1, and 0.75 with DeepSeek-v4. This abstract-level evidence describes the framework and benchmark results but does not provide further evaluation details. ### Source excerpt arXiv:2611.01563v1 Announce Type: new Abstract: The proliferation of Large Language Models has catalyzed powerful application development platforms like Dify, which enable users to build sophisticated LLM-powered workflows through modular node orchestration. However, manually configuring these workflows requires deep platform expertise and remains time-consuming and error-prone, creating a significant barrier for non-expert users. To address this challenge, we propose \textsc{ChatDify}, a novel multi-agent framework that automates end-to-end Dify workflow synthesis from natural language descriptions. \textsc{ChatDify} employs an Actor-Critic architecture comprising three specialized agents: the \textit{Planner Agent} that decomposes user requests into structured execution plans, the \textit{Executor Agent} that performs context-aware parameter synthesis and schema-augmented DSL compilation, and the \textit{Reviewer Agent} that ensures reliability through hybrid verification with closed-loop refinement. We evaluate \textsc{ChatDify} on a curated benchmark of 100 production-inspired tasks across three state-of-the-art LLM backends, achieving 0.96, 0.91, and 0.75 success rates with Qwen-3.7, GLM-5.1, and DeepSeek-v4, respectively. These results demonstrate its robustness and practical applicability for lowering the technical barrier to creating sophisticated LLM applications. ## AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents DevFeed: [AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents]() Original publisher: [Read original article]() Author: Yanjie Li, Yiming Cao, Dong Wang, Bin Xiao Published: 2026-11-03T05:00:00Z Content type: article Language: en Sources: [arXiv: Cryptography and Security]() Topics: [prompt injection](), [LLM security](), [Large Language Model](), [Prompt Engineering](), [Image]() Tags: [abstract](), [adaptive](), [agents](), [art](), [arxiv](), [asr](), [attack](), [attacks](), [attention](), [automatic](), [backends](), [based](), [behavior](), [benchmarks](), [black](), [box](), [but](), [cs-ai](), [cs-cr](), [multimodal](), [prompt](), [prompt-injection]() ### AI overview AgentTypo is a black-box red-teaming framework that embeds optimized text in webpage images to steer multimodal agents toward an attacker's intended answer. It tunes text position, contrast, and color using task-conditioned visual attention, then uses evaluation feedback and successful past examples to refine prompts. On VWA-Adv and Mind2Web, it outperformed the image-based attacks evaluated. Across seven closed- and open-source backends, AgentTypo-pro achieved 45%-83% attack success on VWA-Adv, exceeding CLIP-based AgentAttack by 29-47% and prior typographic attacks by 25-37%. ### Source excerpt arXiv:2510.04257v2 Announce Type: replace Abstract: Multimodal agents built on large vision-language models are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection. Typographic text-in-image attacks such as FigStep have been evaluated mainly on standalone vision-language models. For agents, user goals, webpage context, and action history make the effect of embedded instructions uncertain, and robustness to behavior-guided, adaptive visual injections remains insufficiently studied. We introduce AgentTypo, a black-box red-teaming framework that launches adaptive typographic prompt injection by embedding optimized text into webpage images. We formulate an attention-guided automatic typographic prompt-injection objective that favors the attacker's expected answer over the clean-page answer, together with task-conditioned visual attention from small question-answering models. A novel black-box optimization method searches text position, contrast, and color. To further strengthen the attack, we develop a multi-LLM system that iteratively refines injection prompts using evaluation feedback and retrieves successful past examples for strategy learning. Experiments on the VWA-Adv and Mind2Web benchmarks show that AgentTypo significantly outperforms state-of-the-art image-based prompt injection attacks. Across seven closed- and open-source model backends, AgentTypo-pro achieves 45%-83% ASR, outperforming the CLIP-based AgentAttack by 29-47% and previous typographic attacks by 25%-37% on VWA-Adv. ## Logic Hijacking of RAG: Single-Document Poisoning via Local Optimization DevFeed: [Logic Hijacking of RAG: Single-Document Poisoning via Local Optimization]() Original publisher: [Read original article]() Author: Hangyi Zhao (Faculty of Data Science, City University of Macau, Macau, China), Bonan Zhang (Faculty of Data Science, City University of Macau, Macau, China), Fusen Guo (School of Systems and Computing Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Retrieval Augmented Generation (RAG)](), [Hyperparameter optimization](), [Ground truth / benchmark quality]() Tags: [attack](), [attacks](), [audits](), [automated](), [backends](), [benchmarks](), [black](), [context](), [coupling](), [cs-ai](), [evaluations](), [framework](), [guardrails](), [hijacking](), [llms](), [local](), [logic](), [offline](), [open](), [optimization](), [rag]() ### AI overview MONOLOGUE poisons a RAG corpus through one document by presenting the true ground truth as an authoritative historical baseline, then framing a counterfactual target as a later update. This avoids explicit factual refutation and imperative overrides. Payloads are refined offline using local multi-dimensional gated optimization, combining automated text-quality audits with surrogate behavioral verification to avoid suspicious live writes. The abstract reports the strongest overall attack efficacy among evaluated single-document poisoning attacks across three benchmarks, three retrieval backends, and four open-weight generator regimes; it provides no specific scores or comparisons by backend. ### Source excerpt arXiv:2610.34829v2 Announce Type: replace Abstract: Single-document corpus poisoning under black-box access poses a severe threat to Retrieval-Augmented Generation (RAG). However, existing attacks predominantly rely on explicit factual refutations or imperative instruction overrides. Modern aligned LLMs often arbitrate such conflicts conservatively, rejecting overt contradictions in favor of parametric memory or clean retrieved context while triggering input guardrails. We propose MONOLOGUE, a single-document poisoning framework based on Logic Hijacking without explicit refutation. Rather than disputing the ground truth, our method concedes it as an authoritative historical baseline before introducing the update, weaving a grounded record-correction narrative (e.g., dated institutional reattributions) that presents the counterfactual target as the most recent update. To prevent suspicious live writes, payloads are refined entirely offline via local multi-dimensional gated optimization, coupling automated text quality audits with surrogate behavioral verification. Evaluations across three benchmarks, three retrieval backends, and four open-weight generator regimes show that our method achieves the strongest overall attack efficacy among the evaluated single-document poisoning attacks. ## TIRx: A Unified Tile Primitive ML Compiler Across Heterogeneous Backends DevFeed: [TIRx: A Unified Tile Primitive ML Compiler Across Heterogeneous Backends]() Original publisher: [Read original article]() Author: Bohan Hou, Hongyi Jin, Guanjie Wang, Shushi Hong, Jinqi Chen, Yaxing Cai, Lijie Yang, Zihao Ye, Yaoyao Ding, Ruihang Lai, Todd Mowry, Tianqi Chen Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Distributed Computing](), [arXiv: Machine Learning]() Topics: [Compiler](), [GPU optimization](), [Distributed Systems & Parallel Computing](), [High-Performance Computing]() Tags: [accelerators](), [ai](), [ai-accelerators](), [backends](), [compiler](), [cs-ai](), [cs-dc](), [cs-lg](), [cs-pl](), [dispatch](), [gpus](), [hardware](), [implementation](), [kernels](), [layout](), [layouts](), [memory](), [ml](), [performance](), [scope](), [tensors]() ### AI overview TIRx aims to retain native-level performance control while simplifying ML kernel implementation across GPUs and AI accelerators. It makes participating hardware teams explicit through execution scopes, maps logical tensors to hardware spaces with named-axis layouts and shard/replica/offset semantics, and dispatches tile primitives based on scope, layout, and backend. This automates recurring address calculations, operand preparation, and instruction-loop generation. The abstract reports vendor-level performance on recent GPUs and AI accelerators and substantial reductions in low-level code in native-code case studies; it gives no benchmark details or specific backend comparisons. ### Source excerpt arXiv:2601.19092v3 Announce Type: replace-cross Abstract: Modern ML kernels evolve rapidly with new optimizations and hardware features. ML compiler abstractions must accommodate richer memory hierarchies, more forms of thread cooperation, and increasingly heterogeneous backends while retaining the control required for peak performance. We present TIRx, an ML compiler for GPUs and AI accelerators that exposes these ingredients through explicit interfaces. TIRx introduces an explicit execution-scope abstraction that captures participating hardware teams at multiple granularities. It also uses tensor layouts to map logical tensors to heterogeneous hardware spaces through named axes and shard/replica/offset semantics. Finally, it introduces tile-primitive dispatch, which lowers high-level operations to efficient implementations according to scope, layout, and backend. This design preserves native-level control while automating recurring address calculation, operand preparation, and instruction-loop generation. Our evaluation shows vendor-level performance across the latest GPUs and AI accelerators, while native-code case studies show substantial reductions in low-level implementation code. ## Manual binding and semi-automated interface generation for efficient simulation backends with user-friendly frontends: A~demonstration for epidemic use cases DevFeed: [Manual binding and semi-automated interface generation for efficient simulation backends with user-friendly frontends: A~demonstration for epidemic use cases]() Original publisher: [Read original article]() Author: Maximilian Betz, Daniel-Michel Richter, Daniel Abele, Ren\\'e Schmieding, Henrik Zunker, Martin J. K\\"uhn Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Software Engineering]() Topics: [Model driven development](), [c/c++](), [Simulation]() Tags: [automated](), [backends](), [c-plus-plus](), [code](), [cs-se](), [disease](), [efficiently](), [framework](), [generate](), [generation](), [implementing](), [interactive](), [interface](), [language](), [large](), [libraries](), [model](), [modeling](), [models](), [semi](), [simulation]() ### AI overview The paper presents a way to expose a large C++ infectious-disease simulation framework to Python, combining compiled-backend performance with Python's accessible, iterative modeling workflow. It notes that writing bindings by hand can burden backend developers as models change or expand. To ease that work, it proposes semi-automatically generating interface code by analyzing generic C++ with abstract syntax trees and using modular intermediate representations. The supplied abstract does not report evaluation results for the generation method. ### Source excerpt arXiv:2610.40343v1 Announce Type: new Abstract: Challenges and chances of mathematical software are directly entangled with the programming language used. Depending on the background of developers and users, conflicts between different goals may arise. While several, e.g., scripting languages allow a low barrier and user-friendly entry to mathematical modeling and simulation, corresponding implementations will most likely not allow timely outcomes for detailed simulations. In infectious disease modeling, while several large-scale software frameworks rely on C or C++ implementations, most models are implemented in Python or R. While the latter allow for fast prototyping and more interactive development, it has been shown for several applications that C or C++ implementations outperform plain Python or R implementations by factors between one and three orders of magnitude. In particular, in times of pandemics, the urgent need for swift estimation of the system's dynamics makes the use of large-scale models in these noncompiled languages prohibitive. In order to create synergies and close the gap between scientists from the application and the computational domain, binding libraries allow to execute entire, compiled software frameworks from languages such as Python. However, with iterative model development, extension, or addition of new models, implementing bindings can represent another relevant barrier to developers of the C or C++ backends. In this paper, we show how a large-scale software framework with a multitude of infectious disease models can be efficiently exposed to Python. In addition, we suggest a semi-automated way to generate interface code by analyzing generic C++ code through abstract syntax trees, enabling a modularized approach working with intermediate representations. ## SPARSEBLEED: Bypassing Attribute-Inference Defenses via Sparsity Optimizations DevFeed: [SPARSEBLEED: Bypassing Attribute-Inference Defenses via Sparsity Optimizations]() Original publisher: [Read original article]() Author: Darsh Asher, Farshad Dizani, Joshua Kalyanapu, Ro Cammarota, Aydin Aysu, Samira Mirbagher Ajorpaz Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Cryptography and Security]() Topics: [ai security](), [Neural Network](), [TensorRT]() Tags: [ampere](), [attack](), [avx](), [avx2](), [backends](), [confidence](), [cs-ar](), [cs-cr](), [data](), [deep-neural-networks](), [defense](), [dvfs](), [energy](), [engines](), [execution](), [finance](), [hardware](), [healthcare](), [inference](), [input](), [optimizations](), [sparsity]() ### AI overview SparseBleed infers sensitive attributes from timing signatures produced by sparsity optimizations in neural-network accelerators. The attack uses a single query and, according to the abstract, needs no gradients, logits, weights, labels, power traces, or other side channels. It spans Intel AVX2/AVX-512/AMX, NVIDIA Ampere Tensor Cores with 2:4 sparsity, and nine sparse and dense libraries. On Texas-100X, it inferred Sex with 66.39% accuracy and recovered the Surgical Procedure output label with 99.8% accuracy. Disabling sparsity for constant-time execution carries substantial costs, including up to 2x latency on NVIDIA Sparse Tensor Cores and about 25% more power in AMX microbenchmarks. A worst-case-time padding defense reportedly masked leakage with 7.24% average performance overhead across nine libraries on Texas-100X/Adult and near-zero extra energy. ### Source excerpt arXiv:2610.38110v1 Announce Type: new Abstract: Deep neural networks (DNNs) are widely deployed in safety- and privacy-critical domains such as healthcare and finance, where models are trained on confidential data. This paper presents the first hardware-level attribute-stealing attack that bypasses state-of-the-art privacy defenses, revealing that accelerator optimizations intended for performance can directly undermine data confidentiality. SparseBleed requires no access to gradients, logits, model weights, ground-truth labels, power traces, DVFS manipulation, shared caches, or other channels. Instead, we show that sparsity-driven execution paths encode stable timing signatures that act as a surrogate for model confidence, enabling high-fidelity attribute inference from a single query. SparseBleed generalizes across heterogeneous backends and hardware engines, including Intel AVX2/AVX-512/AMX, NVIDIA Ampere Tensor Cores with 2:4 sparsity, and nine well-known sparse and dense libraries such as Intel oneDNN, oneMKL SparseBLAS, and NVIDIA cuSPARSE. We evaluate SparseBleed to infer sensitive input attributes and true output labels. On Texas-100X, SparseBleed infers sensitive input attributes such as Sex with 66.39% accuracy and recovers the true output label, Surgical Procedure, with 99.8% accuracy. Finally, we evaluate mitigations. Constant-time execution by disabling sparsity optimizations is expensive (e.g., up to $2\times$ higher latency on NVIDIA Sparse Tensor Cores and $\approx 25\%$ higher power in AMX microbenchmarks). By contrast, a padding defense that equalizes responses to the worst-case time masks timing leakage with only 7.24% average performance overhead across nine libraries on Texas-100X/Adult and near-zero extra energy. ## Testing Deep Learning Library APIs via Cross-Framework Differential Fuzzing DevFeed: [Testing Deep Learning Library APIs via Cross-Framework Differential Fuzzing]() Original publisher: [Read original article]() Author: Bin Duan, Ruican Dong, Naipeng Dong, Dan Dongseong Kim, Guowei Yang Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Software Engineering]() Topics: [Fuzzing/Fuzz testing](), [Test automation](), [Deep learning]() Tags: [abstract](), [aliases](), [analysis](), [api](), [apis](), [applications](), [approach](), [arxiv](), [backends](), [boundary](), [candidate](), [cross](), [cs-se](), [deep](), [deep-learning](), [differential-fuzzing](), [framework](), [fuzzing](), [learning](), [library](), [testing]() ### AI overview Xamt tests deep learning library APIs by comparing execution-validated groups of APIs intended to perform equivalent operations across seven libraries. It uses API aliases and parameter-role normalization to propose matches, then checks them pairwise and as groups on ordinary inputs. Variance-guided fuzzing explores ordinary, boundary, and non-finite inputs; crash and output-inconsistency oracles flag cases for reproduction. Xamt reports 676 validated groups spanning 2,563 APIs and 72 independently reproduced discrepancies: 4 crashes and 68 output inconsistencies. Of 72 developer reports, 25 were confirmed and 23 fixed. ### Source excerpt arXiv:2608.11886v2 Announce Type: replace Abstract: Deep learning libraries underpin many safety- and reliability-critical applications, yet existing API-level testing techniques often rely on intra-library properties or CPU--GPU differential oracles and may miss defects that behave consistently across hardware backends. We present Xamt, a cross-framework differential fuzzing approach for deep learning library APIs. Xamt constructs and tests execution-validated groups of APIs intended to implement equivalent operations across seven libraries. It uses explicit API aliases and parameter-role normalization to construct candidate correspondences and validates them through pairwise execution and a group-level behavioral check on canonical ordinary inputs. The resulting groups are explored using variance-guided differential fuzzing with ordinary, boundary, and non-finite inputs. Crash and inconsistency oracles flag executions exhibiting abnormal termination or inconsistent outputs for subsequent reproduction and analysis. Across the seven libraries, Xamt constructs 676 execution-validated groups containing 2,563 matched APIs. Among these, Xamt identifies 72 independently reproduced discrepancy cases, including 4 crash cases and 68 output inconsistencies. Among the 72 developer reports, 25 have been confirmed, including 23 that have been fixed. ## QSoftCore: A Conceptual Foundation for Quantum Software Engineering DevFeed: [QSoftCore: A Conceptual Foundation for Quantum Software Engineering]() Original publisher: [Read original article]() Author: Tao Yue, Yuechen Li, Man Zhang Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Software Engineering]() Topics: [Quantum Computing](), [Software Engineering]() Tags: [abstraction](), [algorithms](), [application](), [applications](), [assessment](), [backends](), [computing](), [concepts](), [cs-se](), [deployment](), [engineering](), [evolution](), [execution](), [foundation](), [lifecycle](), [literature](), [measurement](), [quant-ph](), [quantum](), [quantum-software](), [software](), [software-engineering]() ### AI overview QSoftCore reorganizes established quantum-computing concepts around software engineering, linking applications, problems, algorithms, configurations, programs, and paradigm-specific computation specifications to state evolution, execution results, and backends. The authors report assessing the model against QSE literature across key software engineering activities; they say it provides shared referents for quantum computations, executions, deployment configurations, backends, and measurement evidence across the QSE lifecycle. The supplied text gives no details on the model's structure or assessment methods. ### Source excerpt arXiv:2610.39406v1 Announce Type: new Abstract: Quantum software engineering (QSE) requires understanding application requirements, quantum programs and algorithms, quantum computation paradigms, executions, and deployment. However, relevant quantum computing (QC) concepts are often defined across physical, mathematical, and paradigm-specific levels of abstraction, making it difficult for the QSE community and hence proposed methods to refer to them consistently. To this end, we present QSoftCore, a conceptual model that reorganizes established QC concepts from a software engineering perspective and connects quantum applications, computational problems, algorithms, configurations, programs, and paradigm-specific computation specifications with state evolution, execution results, and backends. Our assessment against QSE literature across key software engineering activities indicates that QSoftCore provides shared QC referents for quantum computations, executions, deployment configurations, backends, and measurement evidence across the QSE lifecycle. ## Characterizing WebGPU Dispatch Overhead for LLM Inference Across Four GPU Vendors, Three Backends, and Three Browsers DevFeed: [Characterizing WebGPU Dispatch Overhead for LLM Inference Across Four GPU Vendors, Three Backends, and Three Browsers]() Original publisher: [Read original article]() Author: J\\k{e}drzej Maczan Published: 2026-10-28T04:00:00Z Content type: article Language: en Sources: [arXiv: Distributed Computing](), [arXiv: Machine Learning]() Topics: [LLM optimization](), [Large Language Model](), [Vercel vgpu](), [RTX PRO]() Tags: [backend](), [backends](), [benchmarks](), [chrome](), [cs-dc](), [cs-lg](), [cs-pf](), [dispatch](), [firefox](), [fusion](), [gpu](), [implementation](), [inference](), [intel](), [linux](), [llm](), [llm-inference](), [macos](), [metal](), [open-source](), [overhead](), [performance](), [webgpu]() ### AI overview At batch size 1, WebGPU dispatch overhead can dominate LLM inference even when kernels are efficient. A sequential-dispatch benchmark across four GPU vendors, two native implementations, three browsers, and two model sizes found that single-operation tests overestimate dispatch cost by about 20x. API overhead alone was 24-36 μs per dispatch on Vulkan and 32-71 μs on Metal; including Python cost, total per-operation overhead was about 95 μs. Backend choice was the main factor, though implementations within a backend also differed (up to 2.2x on Metal). Kernel fusion improved Vulkan throughput by 53%, while CUDA fusion brought no benefit. The authors' torch-webgpu backend reached 11-12% of CUDA performance on their reference platform. These results cover the tested batch size, models, platforms, and pipeline; code, benchmarks, and raw data are open source. ### Source excerpt arXiv:2604.02344v2 Announce Type: replace Abstract: WebGPU's security-focused design imposes per-operation validation that compounds across the many small dispatches in neural network inference, yet the true cost of this overhead is poorly characterized. We present a systematic characterization of WebGPU dispatch overhead for LLM inference at batch size 1, spanning four GPU vendors (NVIDIA, AMD, Apple, Intel), two native implementations (Dawn, wgpu-native) and three browsers (Chrome, Safari, Firefox), and two model sizes (Qwen2.5-0.5B and 1.5B). Our primary contribution is a sequential-dispatch methodology that reveals naive single-operation benchmarks overestimate dispatch cost by ${\sim}20\times$. The true per-dispatch cost of WebGPU API overhead alone is 24-36 $\mu$s on Vulkan and 32-71 $\mu$s on Metal, while the total per-operation overhead including Python cost is ${\sim}95$~$\mu$s, which turns out to be a distinction critical for optimization. On Vulkan, kernel fusion improves throughput by 53%, while CUDA fusion provides no benefit, confirming that per-operation overhead is a primary differentiator. LLM inference was tested across three major operating systems (Linux, Windows, macOS). We built $\texttt{torch-webgpu}$, a PrivateUse1-based out-of-tree PyTorch backend and an FX-to-WebGPU compiler, which on our reference platform achieves 11--12% of CUDA performance. At dtype-matched float32, RTX PRO 2000 achieves 1.4$\times$ WebGPU's throughput despite ${\sim}6\times$ less compute than RTX 5090. For dispatch overhead, backend choice is the dominant factor, although implementation choice also matters substantially within a backend (2.2$\times$ for Metal). In terms of dispatch vs kernel compute efficiency, we conclude that at batch=1 with the current dispatch-heavy pipeline, per-operation overhead dominates regardless of kernel quality. All code, benchmarks, and raw data are open source. ## The Agent in the Middle: How AI-Mediated Access Reshapes Web Vulnerability Exploitation DevFeed: [The Agent in the Middle: How AI-Mediated Access Reshapes Web Vulnerability Exploitation]() Original publisher: [Read original article]() Author: Ruozhao Yang, Jiongchi Yu, Ming Hu, Xiaofei Xie Published: 2026-10-27T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Cryptography and Security](), [arXiv: Software Engineering]() Topics: [Security](), [AI Agent](), [AI Bots]() Tags: [abstract](), [account](), [agent](), [agents](), [ai](), [ai-agents](), [also](), [application](), [architecture](), [arxiv](), [attack](), [attacks](), [backend](), [backends](), [behavior](), [benchmark](), [but](), [cs-ai](), [cs-cr](), [cs-se](), [exploitation](), [vulnerability](), [web]() ### AI overview VELA evaluates 44 Web vulnerabilities across 12 application families and finds that agent-facing interface design mainly determines whether existing attacks remain exploitable. In the reference configuration, 91% remain exploitable, and agents realize every attack that direct interface invocation achieves. Restrictions that remove operations or limit input control reduce exploitability, but tested interfaces span 39% to 95% attack retention; interface choice shifts retention more than agent choice (at most 4.5 percentage points across agents). The study attributes no reference-configuration losses to refusal or input normalization, and reports one case where an agent supplied omitted attack calls. Restrictive interfaces can also limit legitimate workflows, exposing a security-functionality trade-off. ### Source excerpt arXiv:2610.35541v1 Announce Type: new Abstract: Web services increasingly expose backend interfaces that let users delegate tasks to AI agents, which invoke service operations directly without navigating websites. This shift changes how existing Web vulnerabilities can be exploited. We investigate these changes through VELA, an executable benchmark covering 44 vulnerabilities across twelve Web application families. The benchmark preserves vulnerable backends and derives agent-facing interfaces from legitimate workflows. We evaluate three commercial agents across multiple interface configurations, using controlled comparisons to separate interface effects from agent behavior and independent oracles to verify attack outcomes. Our results show that the agent realizes every attack that direct interface invocation realizes in the reference configuration, so exploitability is set mainly by the interface: 91% of vulnerabilities remain exploitable under the reference design. Interface restrictions prevent some attacks by removing required operations or limiting input control, while retention remains high across vulnerability classes. No case-level loss in the reference configuration is attributed to agent refusal or input normalization, and in one case the agent completed an attack by supplying calls the attacker omitted. Across the tested configurations, interface design affects exploitability more than agent choice, moving retention from 39% to 95% against at most 4.5 points across agents, but some restrictive interfaces also limit legitimate workflows. Our study provides a systematic empirical account of how agent-mediated access reshapes existing Web vulnerabilities, grounding future Web architecture design in measured security outcomes and trade-offs between reducing attack exposure and preserving legitimate functionality in agent-facing services. ## Context Is Not Free: Carbon-Aware LLM Routing for Multi-Turn LLM Web Services across Geo-Distributed Data Centers DevFeed: [Context Is Not Free: Carbon-Aware LLM Routing for Multi-Turn LLM Web Services across Geo-Distributed Data Centers]() Original publisher: [Read original article]() Author: Xiaoyang Zhang, Yuejun Yan, Qingang Zhang, Yonggang Wen Published: 2026-10-27T04:00:00Z Content type: news Language: en Sources: [arXiv: Distributed Computing]() Topics: [Model Routing](), [LLM optimization](), [Large Language Model](), [data centers](), [AI Context Management]() Tags: [backends](), [cache](), [carbon](), [carbon-footprint](), [context](), [cost](), [cs-ai](), [cs-cy](), [cs-dc](), [data-centers](), [datasets](), [deployment](), [distributed](), [emissions](), [experiments](), [framework](), [gate](), [gpu](), [llm](), [routing](), [services](), [web](), [web-services]() ### AI overview CarbonRouter reduces emissions in multi-turn LLM services by accounting for the carbon cost of rebuilding conversation history when a session moves between deployments. Because each session's KV cache is local to the deployment that served earlier turns, recomputation grows with conversation length and can outweigh the benefit of moving to a cleaner region or smaller model. The online framework jointly chooses model, GPU type, and region, estimates recomputation using cached-token counts from serving backends, and uses a lightweight gate to avoid moves likely to be reversed on the next turn. In trace-driven experiments on two conversation datasets, five LLMs, three GPU types, and four regions, it reduced CO₂ by 16.1% on average versus FairCarbon, which has the same action space, while improving predicted answer quality and estimated latency. The reported reduction was 23.1% for sessions longer than 30 turns; these are experimental results, not evidence of production deployment. ### Source excerpt arXiv:2610.35230v1 Announce Type: new Abstract: The sustainability of the Web has drawn growing attention, as large language model (LLM)-based web services leave a substantial carbon footprint. These services typically run in geo-distributed data centers whose grid carbon intensity varies across regions and over time, so routing each query to a cleaner region or a smaller model can reduce emissions. A conversation, however, is a stateful web session: each turn reuses the key--value (KV) cache of its history, which resides only where earlier turns were served. Serving a turn on another deployment requires recomputing this history, whose carbon cost grows with the conversation and can outweigh the saving of the move. Existing carbon-aware LLM routers treat a conversation's queries as independent and ignore this cost. We present CarbonRouter, an online framework that jointly selects the model, GPU type, and region for each query to minimize carbon emissions. CarbonRouter introduces a cache-aware carbon cost model, which charges history recomputation to every candidate deployment using cached-token counts from the serving backends, and uses a lightweight gate that keeps a session on its cached deployment when it would likely move back at the next turn and rebuild its history again. We evaluate CarbonRouter in trace-driven experiments on two real-world conversation datasets with five LLMs, three GPU types, and four regions, where it achieves the lowest carbon of all policies. CarbonRouter reduces CO$_2$ by 16.1\% on average relative to FairCarbon, a state-of-the-art carbon-aware router with the same action space, while also improving predicted answer quality and estimated latency. Moreover, the saving rises further to 23.1\% for sessions longer than 30 turns. ## Introducing Temporal Proxy DevFeed: [Introducing Temporal Proxy]() Original publisher: [Read original article]() Author: Luke Knepper Published: 2026-10-27T00:00:00Z Content type: release Language: en Sources: [Temporal]() Topics: [Secrets Management](), [sensitive data](), [payload](), [Cloud]() Tags: [agents](), [ai-agents](), [application](), [applications](), [aws](), [aws-kms](), [azure](), [backends](), [cloud](), [control](), [custom](), [data](), [encryption](), [keys](), [libraries](), [payload](), [product-news](), [proxy](), [sensitive-data](), [teams](), [temporal]() ### AI overview Temporal Proxy centralizes encryption of Temporal Workflow payloads in a component teams deploy and operate in their own environment. Applications connect to it over authenticated mTLS; it encrypts payloads with short-lived AES-256-GCM data keys protected by customer-controlled keys through AWS KMS, Google Cloud KMS, Azure Key Vault, or custom backends. Temporal Cloud stores ciphertext and protected data keys but cannot decipher them; the proxy decrypts values for Workers within the customer environment. To reduce accidental bypass, the text recommends making the proxy the required connection, keeping Cloud credentials with it, and blocking direct application and Worker connections. Configure encryption for application payloads, error messages, and stack traces; Search Attributes remain unencrypted, and existing stored data is not encrypted retroactively. The text says the product is in Public Preview, fully supported, and recommended for production. ### Source excerpt Temporal Proxy centralizes payload encryption with keys you control, so Temporal Cloud can't read sensitive data and teams skip custom encryption libraries. ## The next 10 years of KubeVirt DevFeed: [The next 10 years of KubeVirt]() Original publisher: [Read original article]() Author: Vladik Romanovsky Published: 2026-10-26T14:00:00Z Content type: article Language: en Sources: [Kubernetes Overview, News and Trends | The New Stack](), [The New Stack]() Topics: [virtualization](), [virtual machines](), [AI Infrastructure](), [trust]() Tags: [2016](), [abstraction](), [accelerated](), [ai](), [ai-cloud](), [ai-inference](), [ai-workloads](), [amd](), [api](), [arm](), [attestation](), [backend](), [backends](), [bare-metal](), [beta](), [blackwell](), [cloud-native-ecosystem](), [cncf](), [containers](), [kubernetes](), [open-source](), [post-contributed](), [red-hat](), [sponsor-cncf](), [sponsored-post-contributed](), [virtual-machines]() ### AI overview KubeVirt is evolving from a way to run legacy VMs alongside containers into a broader Kubernetes platform for varied virtualization workloads. Its roadmap favors Kubernetes-native resource allocation: GPU claims are beta in v1.9, with coordinated claims for other devices and CPUs in development. Confidential VM support spans multiple hardware platforms but varies in maturity: IBM Secure Execution is GA, AMD SEV/SEV-SNP is beta, Intel TDX is alpha, and Arm CCA remains under development. The project is also developing structured plugins for features outside core (alpha in v1.9) and a pluggable virtualization framework: MSHV support is alpha, while broader support for out-of-tree stacks such as Cloud Hypervisor, OpenVMM, and Firecracker is ongoing. Attestation and protected accelerator communication remain exploratory; future capabilities depend on review, integration, and testing. ### Source excerpt KubeVirt was founded by Red Hat in 2016 around the idea that Kubernetes could run virtual machines alongside containers. We The post The next 10 years of KubeVirt appeared first on The New Stack. ## Node.js Brotli Decoder Out-of-Bounds Write (RCE) - CVE-2026-111497 DevFeed: [Node.js Brotli Decoder Out-of-Bounds Write (RCE) - CVE-2026-111497]() Original publisher: [Read original article]() Author: Nir Zadok Published: 2026-10-22T14:50:56Z Content type: news Language: en Sources: [OX Security]() Topics: [Memory Safety](), [cve remediation](), [декодер]() Tags: [2026](), [apis](), [backends](), [block](), [buffer](), [bug](), [build](), [built-in](), [code](), [cve](), [data](), [decoder](), [http](), [memory](), [node-js](), [october](), [process](), [releases](), [research](), [responses](), [security](), [write]() ### AI overview Node.js releases on October 21, 2026 fixed a high-severity out-of-bounds write in the bundled Brotli decoder. Decompressing attacker-controlled Brotli data--including fetched HTTP responses encoded with `Content-Encoding: br`--could cause memory writes past the buffer, potentially leaking memory or enabling code execution with the process's privileges. The bug arose when malformed block switches were treated as recoverable or their errors discarded, letting decoding continue with invalid state. Upgrade to Node.js 22.23.4, 24.22.1, or 26.12.1; end-of-life lines 25.x, 23.x, and versions 20.x or earlier receive no fix and should move to a supported line. The excerpt identifies direct and indirect Brotli decompression of untrusted input as the sensitive path and advises relying on official releases rather than unverified exploit or patch scripts. ### Source excerpt On October 21, 2026, Node.js shipped security releases across all supported lines fixing a high-severity out-of-bounds write in the bundled Brotli decoder. A Node.js process that decompresses attacker-controlled Brotli data -- including Brotli-encoded HTTP responses it fetches -- can be driven to write past the end of its buffer, leaking memory or executing code with [...] The post Node.js Brotli Decoder Out-of-Bounds Write (RCE) - CVE-2026-111497 appeared first on OX Security. ## Google turns the infrastructure behind Gemini's agents into a cloud service DevFeed: [Google turns the infrastructure behind Gemini's agents into a cloud service]() Original publisher: [Read original article]() Author: Frederic Lardinois Published: 2026-10-22T13:00:00Z Content type: news Language: en Sources: [The New Stack]() Topics: [Google Cloud Platform (GCP)](), [Gke](), [Infrastructure as a Service (IaaS)](), [Container as a Service (CaaS)](), [openstack]() Tags: [access-control](), [agent](), [agent-compute-engine](), [agent-gateway](), [agent-sandbox](), [agentcore](), [agentic](), [agents](), [ai](), [ai-agent](), [ai-agents](), [ai-infrastructure](), [aws](), [backends](), [bedrock](), [cloud](), [gemini](), [google](), [google-cloud](), [infrastructure](), [preview](), [service]() ### AI overview Google plans to offer Agent Compute Engine in preview to Google Cloud users later this year. It gives each agent a persistent, hardware-isolated microVM with root access, checkpointing idle sessions so they consume no active resources while waiting. Security design assumes an attacker controls the sandbox: sessions have no stored secrets, use short-lived identities, and reach services through an external proxy that checks delegation before attaching credentials; traffic and stored snapshots are encrypted. The service runs on Cloud Hypervisor and Google's Titan and Titanium IPU platform. For teams managing their own infrastructure, Agent Substrate is an open source GKE runtime that supports microVM or gVisor isolation and runs on any Kubernetes cluster. Google reports up to 10 times the density of traditional compute and, in a 200,000-agent benchmark, 500 activations per second with suspend and resume under half a second. These figures are company claims. ### Source excerpt Google will soon make the infrastructure that powers its Gemini agents available to Google Cloud users in preview. Announced on The post Google turns the infrastructure behind Gemini's agents into a cloud service appeared first on The New Stack. [Next page]()