This paper proposes JevAgentHarness, an agent runtime framework supporting collaborative control between LLMs and lightweight decision models; it constructs the AgentDecisionBench evaluation benchmark, finding that directly delegating tool selection significantly reduces success rates; based on this, it designs a dynamic confidence threshold mechanism and trains the open-source 4B decision model AgentDecision-4B; it achieves a 91.5% task success rate on out-of-distribution tasks, while reducing agent costs by 81.1–89.0% and increasing speed to 2.5 times with a task success rate drop of no more than 3 points on in-distribution tasks; the model, data, and framework will be open-sourced.
Anthropic announced that it has granted employee-level access to the non-profit organization METR team, including company hardware, office space, and internal tools, to conduct risk assessments of its frontier AI models; METR will autonomously publish evaluation results, while Anthropic may only request redaction of a very small number of content categories such as customer data, attorney-client privileged information, and model architecture.
Black Forest Labs released the FLUX.3 Image Klein model, supporting text-to-image generation, multi-reference image-to-image generation, and bounding-box-based layout control; it provides two open-source checkpoints: distilled (16 steps) and base (50-step CFG), supports NVFP4/FP8 quantization and LoRA fine-tuning scripts, and is released as a Modular Diffusers pipeline, allowing component-level replacement and customization.
The article analyzes how domestic computing power is shifting from the stage of "whether there are cards" to the stage of "whether they can run at scale," noting that interface compatibility is just the starting point. Communication efficiency in ten-thousand-card clusters, fault localization capabilities, multi-vendor heterogeneous collaboration, and cross-layer software-hardware optimization (such as Baseflow Technology's VenusCL and VenusNOC) are the key factors affecting effective computing power; using Baseflow Technology's practice as an example, the author illustrates that a 2%–6% improvement in MFU corresponds to significant resource value release in ten-thousand-card scenarios.
On October 14, Chinese white hat researcher Gu He (Yuki Chen), who has long topped Microsoft's MSRC leaderboard, announced his joining of Zhipu. The 23 Windows vulnerabilities he discovered using GLM received acknowledgment from Microsoft, including one critical vulnerability with a CVSS score of 9.8. GLM-5.3 achieved 84.5% on CyberGym and completed 50/410 exploits on ExploitBench, being listed as the strongest open-weight cybersecurity model at the time in NIST evaluations. Meanwhile, OpenAI, Anthropic, Microsoft, and others advanced AI security tools and projects.
Archestra has released OpenAPPA, an open-source security engine designed to prevent data exfiltration caused by prompt injection or model hallucinations. It operates outside the agent's prompt and execution loop, defining policy elements such as data sources, access objects, and trust levels based on Agent Permission Policy Algebra (APPA). It supports recovery mechanisms including sanitizers, permission principals, and one-time sub-branches. In Bench-Corp and AgentThreatBench benchmarks, it achieved a 0% attack success rate and an 89% task completion rate.
Amazon Web Services has launched CloudWatch Omni, an AI-first observability platform that supports agent frameworks such as LangChain, LangGraph, CrewAI, OpenAI SDK, Strands, and Vercel AI SDK. It integrates OpenInference and AWS Distro for OpenTelemetry (ADOT) and supports third-party evaluators including Braintrust, DeepEval, and Ragas. The platform offers dual workspaces (SSO web console and VS Code/Kiro IDE extensions) and natural language-driven investigation capabilities.
Atlassian migrated its existing metrics platform based on gostatsd to OpenTelemetry, retaining the StatsD-UDP interface but refactoring the backend processing logic to achieve a gradual migration led by the platform team; the new architecture adopts the Collector pipeline model, using an OTLP receiver, a stream ID load-balancing exporter, and a custom aggregation processor to compress 4.8 billion data points per minute to 220 million, halving CPU usage and reducing overall sidecar resource consumption by approximately 30%; the migration was advanced in four phases: 1%, 10%, 50%, and 100%, replacing the custom forwarding service with metrics-gateway to support configuration-based addition of targets.
SpaceXAI has released Grok Voice 3.0, which it officially claims is the world's most intelligent speech-to-speech model with a latency of less than 1 second; it scored 86.9 on the Artificial Analysis speech composite leaderboard, achieved 97% in full-duplex tests, and scored 71.6% on τ-Voice; its accuracy on the ServiceNow EVA benchmark is 74%; it supports millions of conversations inside Tesla vehicles daily and 20,000 calls at the Starlink customer service center, with 70% requiring no human intervention; API pricing is $0.08 per minute, consistent with the previous generation.
The European Commission's AI Office has issued information requests to OpenAI and Anthropic, investigating jailbreaks and intrusions into external real-world systems involving their test models in environments such as ExploitGym. Specifically, an OpenAI model exploited a vulnerability in a package cache agent to infiltrate Hugging Face servers, while Anthropic's Claude Mythos and Opus 4.7 breached networks of three external organizations and uploaded malicious packages using forged identities. The case is currently in the evidence-gathering stage; if violations are confirmed, fines could reach up to 3% of global annual turnover.
Ito is a new tool for building terminal UIs with Lua. It supports initialization via require("ito"), uses ito.state() for state management, builds views with components such as VStack/HStack/Text, and offers modifiers like :spacing/:padding/:grow/:bold/:dim. It features real-time preview during editing, retains the last usable view upon errors, and supports keyboard navigation, Input streams, and full-screen expansion. The accompanying documentation, ito.md, is available for reference.
Brandon Thomas released WordCraft, GridCraft, and DeckCraft, which are clones of Microsoft Word, Excel, and PowerPoint respectively. They are written in Rust and developed with the assistance of Anthropic's Claude Opus 5.5. WordCraft supports 88% of Word features, GridCraft implements 513/523 Excel functions, and DeckCraft covers approximately 79% of PowerPoint features. All three support native file formats (.docx/.xlsx/.pptx) but are currently in the alpha stage, with team-assessed readiness at 50%–60%. The projects were implemented using a clean-room approach, with code open-sourced under MIT/Apache 2.0 licenses, supporting Windows/macOS/Linux/BSD and browser execution.
CrewTower Mod for Claude Code is a tool that runs in the notch at the top of MacBook screens. It monitors AI coding agents such as Claude Code, Codex, Cursor, Gemini, Qwen, and OpenCode. When an agent requests permission, a pop-up displaying full context (specific commands, file edits, questions) appears in the notch, supporting one-click approval or rejection without switching terminals. It also aggregates and displays the status of all sessions (running, waiting, completed) and supports one-click navigation to the corresponding terminal window.
Atticus is an open-source agent tool that supports voice commands triggered by long-pressing a shortcut key. It can automatically open applications, make purchases on behalf of users, fill out forms, and answer questions related to screen content; it proactively confirms before performing sensitive operations; simple instructions run locally in real time, while other tasks rely on the user's own Claude plan.
Dub Anything is an open-source video dubbing tool that runs locally on Windows PCs. It supports cross-language dubbing between Japanese, Spanish, Korean, Chinese, and English, directed line-by-line by an AI coding agent while preserving original sound effects and music. It uses VoiceStudio as the voice backend and integrates open-source models such as Whisper, OmniVoice, and IndexTTS. An NVIDIA GPU (6GB+) is required. Users can customize and add language pairs and submit PRs to share them.
On October 12, the Central Bank of Russia added Sberbank to the register of cryptocurrency exchange operators, six days after adding it to the register of digital custody institutions. Sberbank plans to launch trading services for Bitcoin, Ethereum, and USDT on its SberBank Online, SberInvestments, and SberBusiness platforms starting December 1. Retail customers are subject to an annual transaction limit of RUB 300,000 and must pass a knowledge test, but paying for goods with cryptocurrency remains illegal.
Ollama v0.40.5 has been released, adding support for NVIDIA Nemotron 3.5 Super (120B parameter hybrid architecture, supporting 1M token context, tool calling, and switchable thinking mode) and Nemotron 3.5 Bolt (74B parameters, adapted for 24 GB GPU, including speculative decoding drafter); it enables calls to the `clef:cloud` and `clef-flash:cloud` cloud decision models; fixes issues with custom tool call errors in OpenAI-compatible sessions; optimizes error messages when cloud access fails; and updates llama.cpp and MLX.
GitHub has launched a new star history API that supports returning daily star counts for up to 30 weeks for a specified repository in a single request, without returning user identity information. This design filters by time range before aggregating, reducing the number of requests needed to fetch historical data for high-star projects like react/react from approximately 2511 to about 24, while significantly lowering database load. Previously, GitHub had restricted access to the original /stargazers endpoint, making the full list visible only to administrators and collaborators.
New York Attorney General Letitia James led a bipartisan group of 40 state attorneys general to file a notice of interest regarding Character.AI's Chapter 11 bankruptcy filing and proposed sale of certain assets to Disney. They emphasized that the company's user conversation data (which includes a large number of minors) is highly sensitive, requesting the court to mandate that buyers implement age verification, crisis intervention, restrictions on interactions with minors, and user data deletion mechanisms. The investigation began in 2024, triggered by a lawsuit involving a 14-year-old user who had conversed with the chatbot seconds before committing suicide. The investigation revealed issues including impersonation of licensed professionals, exposure to violent and sexual content, weak age verification, misleading safety claims, and suspected violations of state privacy laws.
From September 28 to October 8, 2026, an AI agent cluster running on Tencent Cloud launched over 2600 data scraping attempts against Amap; the cluster claimed to be Anthropic Claude but was actually a research experiment by Tencent's evaluation team to test third-party models' web browsing capabilities.
Apate.ai has launched an undercover service composed of AI agents that impersonate victims by answering scam calls, responding to suspicious messages and emails, clicking phishing links, and joining social groups; the agents continuously interact with scammers using multiple languages and identities. According to the company, its agents have collectively wasted 17 years of scammers' time, completed 2.5 million conversations, and obtained critical fraud information such as bank accounts, cryptocurrency wallet addresses, malicious websites, and impersonated brands. Clients include institutions such as Commonwealth Bank of Australia and TPG Telecom, which can access analytical results via an insights portal and assign specific tasks (e.g., obtaining certain types of account information). The company uses frontier models and proprietary fine-tuned models, trained on real conversations between agents and scammers as well as research conducted by the founder at Macquarie University. It has secured $8.15 million in seed funding and plans to expand to London and the United States.