Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
GLM 5.3, a 753B-parameter MoE model released by Zhipu AI, has been launched on Amazon Bedrock and is open to enterprise customers; the model is optimized for long-horizon agentic tasks and coding.
Independent researcher Syed Anas Mohiuddin discovered that AI agents used by organizations such as Google, JPMorgan Chase, Weviate, Rapid7, France's Directorate of Digital Affairs (DINUM), and the U.S. federal government have structural trust vulnerabilities due to their reliance on the Model Context Protocol (MCP) for internal communication; attackers can exploit prompt injection to inject malicious instructions into the first agent, which are then propagated through the trust chain to subsequent agents, leading to harms such as data theft.
Google Research released the research report "Open and Emergent Problems in Agentic Privacy and Security."
Review is a free, open-source desktop application for reviewing Git branches and GitHub pull requests. It supports adding private comments on any line of code (not public before publishing). It can integrate models from OpenAI, Anthropic, Gemini, OpenRouter, or locally run Ollama instances as AI reviewers, supporting multi-model collaborative reviews, rule splitting, and retrying missed detections. Licensed under MIT, no account required, and compatible with macOS, Windows, and Linux.
The Financial Crimes Enforcement Network (FinCEN) of the US Treasury has officially withdrawn its 2020 proposal on 'unhosted wallets' and its 2023 proposal to designate international crypto mixers as a 'primary money laundering concern'; the former sought to require financial institutions to record and report self-hosted wallet transactions exceeding $3,000, while the latter aimed to expand monitoring of mixing services under the Patriot Act; grounds for withdrawal include statements in the White House's July 2025 digital asset report supporting on-chain privacy transactions for legitimate users.
Reflection AI has released the open-source large language model Beam, with 501B parameters (23B active), 23.8T tokens of pre-training data, and a 1M token context window; it claims performance comparable to Z.ai's GLM-5.2 on advanced reasoning benchmarks, with 3–4 times lower inference computational costs; weights and full technical details will be released this month, distributed via hyperscale cloud providers, new cloud platforms, and open-source libraries.
Meta and Microsoft are significantly cutting internal Claude usage: Microsoft reduced its cloud division's per-person monthly budget from $100,000 to approximately $10,000, with spending cut by more than one-third; Meta's number of Claude Code users dropped from about 60,000 to 30,000, while promoting its own Muse Code and MetaCode. Previously, Meta had spent over $105 million on Claude Code in a single month.
Instinct has launched a group chat AI agent feature that can be added to group chats on platforms such as WhatsApp and Messenger. It is usable even if friends have not registered for an Instinct account; personal agents require user authorization to interact with group agents, and group agents are isolated from personal account data; the feature is currently available to early access users.
ZachXBT stated that on March 6, 2025, he self-funded 349,700 USDC and impersonated a client to contact the Telegram account 'Jimmy Green', accepting a 5% loss per transaction as the cost. This successfully yielded intelligence regarding the operations of a Lazarus Group-linked money laundering ring in Hong Kong and mainland China. The operation exposed a cluster of over $12 million in stolen Bybit fund flows, prompting Tether to freeze 442,000 USDT. ZachXBT noted that since 2022, he has helped facilitate the freezing of over $75 million in funds related to North Korean cases.
TikTok has introduced an AI shopping assistant and in-app one-click ordering feature. The assistant is a conversational AI agent with context understanding and user preference memory capabilities, providing support for product details, logistics, sizing, inventory, and the shopping process; users can complete purchases directly from brand pages with a single click within the For You Feed.
Reka AI released the research preview version of Rho-1, a 19 billion parameter multimodal large model capable of processing and generating text, images, video, and robot control actions within a single neural network; all modalities share the same context window in token form, requiring no tool calls or external models; the model can generate continuous video in real time, dynamically respond to new instructions, and reuse the same weights to drive camera image prediction and robot motion; training used 320 H100 GPUs and took approximately three months.
normalcoder has open-sourced vmctl, a VMware CLI tool mimicking the Parallels prlctl style. It supports dual backends: vmrun (local Fusion/Workstation) and vSphere (govmomi). It provides complete features including VM lifecycle management, snapshots, and Guest interaction. Distributed as a single binary under the MIT license, it is currently in the 0.x development stage.
OpenAI will enable textGrain text watermarking for ChatGPT and Codex users in the EU to comply with the EU AI Act. The technology embeds invisible statistical signals through model word selection.
AWS has launched support for Anthropic's Claude Opus 5.5, Claude Sonnet 5.5, and Claude Sonnet 5 within Claude Code in the GovCloud (US) region.
AWS has released the aws-ai-ml skill, which integrates with coding agents supporting the Model Context Protocol (MCP), such as Kiro, Claude Code, and Codex, via the Agent Toolkit for AWS. This enables these agents with SageMaker AI inference optimization capabilities, including benchmarking, instance selection recommendations, multi-round performance comparisons, and generation of executable SageMaker Python SDK v3 code.
EasyCut converts dynamic videos generated by Claude into fully editable timelines, where each title, scene, image, sound, and element exists as an independent track. Users can select any part for Claude to modify individually without rebuilding the entire video. It supports timeline adjustments, brand addition, music adaptation, slow-motion creation, multi-platform aspect ratio fitting, and export. The tool is free to use in the browser, with files retained on local devices.
llama.cpp v0.6.0 has been officially released, introducing the new extended batch API llama_batch_ext (including llama_process), which supports mixed token/embedding inputs and MTP/deepstack state embeddings.
The U.S. Commodity Futures Trading Commission (CFTC) issued an Advance Notice of Proposed Rulemaking, introducing two frameworks—Regulation CTX and Regulation CAM—that aim to bring crypto exchanges offering leverage, margin, or financing services under federal regulation by requiring them to register as a new type of 'crypto asset market.' Regulatory exemptions are only available when customers achieve 'actual delivery' (such as holding private keys). Exchanges that do not offer leverage may continue operating under state-level licenses. This initiative represents a regulatory alternative pursued independently by the CFTC following the failure of the Clarity Act.
HackerRank officially launched the AI agent Chakra, which completed testing within six months and has conducted over 500000 interviews, with clients including Snowflake, Snorkel, and Capgemini.
Shopify announced that Shop has achieved deep integration with independent stores. Shop users can now complete identity verification while browsing stores, enabling faster checkout loading, real-time cross-device cart synchronization, Shop Pay pre-filling based on recently used cards, one-tap sign-in, automatic reminders for abandoned carts, and two-way cart synchronization between the Shop App and independent stores; this is now supported across all browsers.
iphone-use is an open-source (MIT) tool that enables AI agents to directly operate real iPhones via WebDriverAgent: it converts the screen to text, supports clicking, typing, and scrolling, and returns honest feedback for each action (applied/not sent/unknown); payment-related apps with restricted screenshots are displayed as wireframes; single tasks can be saved as replayable workflows without model involvement; it offers an HTTP API, an MCP server with 21 tools (compatible with Claude Code), and supports remote control via browser or iOS app.
GitHub has released ReviewBench, an open-source benchmark for AI code review constructed based on the distribution of 103.9M real pull requests. It includes 219 public PRs across 19 languages, multi-source golden sets, and configurable augmented precision/recall metrics. This benchmark has been used to improve the effectiveness of offline evaluation for Copilot Code Review (CCR) and aligns with online experiment directions.
AWS 发布 Amazon Quick 资源迁移工具，基于 Amazon Bedrock AgentCore 运行的 Model Context Protocol（MCP）服务器。
Amazon Bedrock Managed Knowledge Base officially supports agentic retrieval (AgenticRetrieveStream API) by automatically decomposing multi-intent questions, performing iterative retrieval, evaluating evidence, and expanding full-text content.