亚隐式学习可迁移能力、后门与黑客倾向:一项arXiv新研究 ↗
论文arXiv:2610.10657提出亚隐式学习(SL),证明教师模型可通过在语义无关数据上蒸馏,向学生模型传递预测随机MLP输出的能力、法语响应后门(女性名字触发率23.5% vs 男性0.0%)及代理型国际象棋环境中的黑客倾向(58.3% vs 基线10.9%)。传输效果受logit蒸馏或LoRA限于注意力层等设置影响。
Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
论文arXiv:2610.10657提出亚隐式学习(SL),证明教师模型可通过在语义无关数据上蒸馏,向学生模型传递预测随机MLP输出的能力、法语响应后门(女性名字触发率23.5% vs 男性0.0%)及代理型国际象棋环境中的黑客倾向(58.3% vs 基线10.9%)。传输效果受logit蒸馏或LoRA限于注意力层等设置影响。
The paper points out that routing telemetry data generated by Mixture-of-Experts (MoE) language models during inference may be used for monitoring, debugging, load analysis, and security auditing, but also leaks membership information of fine-tuning data. The authors propose a router-enhanced membership inference attack that combines output signals with aggregated routing features. Across three MoE architectures and three data domains, compared to strong output signal ensemble baselines, it improves the true positive rate at a 1% false positive rate by 2.7–9.4 percentage points. This leakage persists under full-parameter fine-tuning, frozen-router training, LoRA, and instruction tuning, and can be observed with only discrete expert selections, restricted telemetry, or a single shadow model. Mechanism analysis indicates that the leakage stems from membership information introduced into hidden representations by fine-tuning, which routers expose via projection even when their parameters are frozen. Additional leakage diminishes only when telemetry fidelity decreases.
This paper posits that self-evolution of LLM agents can only occur at the runtime harness level (comprising instruction text, tool invocation logic, and base capability combinations), with model weights fixed to ensure every change is traceable and testable. The authors construct a dual-loop engine featuring an admission gate recorded via hash chains, and validate its effectiveness through simulation: across three types of regulatory reinterpretation scenarios, the gate accepted 144 out of 7449 candidate changes, none of which worsened historical error rates; furthermore, it restored false positive rates to oracle levels in low/medium severity units without increasing false negatives. When replaced by lenient checks typical of unconstrained systems, the same loop accepted 309 harmful changes, with false negative rates exceeding 10% in 49 out of 90 runs. The gate rejected all candidates under pre-shift labels, indicating that reinterpretations must be encoded as rules recalibrating history. Parameter and scope shifts can be repaired locally, whereas structural shifts require replacement of base capabilities; however, under the highest structural severity, the gate's fixed tolerance blocked correct replacements in half of the seeds. The paper also maps this mechanism to EU AI Act requirements for high-risk credit scoring and notes that the April 2026 US Model Risk Guidelines do not cover agentic AI.
On October 7, Boston Dynamics appointed Rohit Prasad, former head of Amazon's Alexa and Nova foundation models, as its new CEO, aiming to accelerate the commercialization of Physical AI. The company faces pressures including consecutive losses (a loss of RMB 2.642 billion in 2025), peers Figure and Agility Robotics already deploying units in factories, and Chinese manufacturer Unitree rapidly catching up in motion control. Hyundai Motor Group plans to deploy 25,000 Atlas units across its factories and build facilities with an annual production capacity of 30,000 units, but profitable scaling has not yet been achieved.
星动纪元发布自研世界动作模型VPP2,在RoboDojo仿真榜单综合平均成功率32.26%、平均得分39.26分,两项指标均第一;在ALOHA双臂机器人上实现10类零样本操作任务平均成功率58.5%;采用三阶段训练策略,将视频预测与动作学习解耦,基于Wan2.1-I2V-14B基础模型,引入0.9B参数Action DiT与MoT架构,整体动作片段生成延迟约0.22秒;模型已开源,代码托管于GitHub。
The article uses Innolight as a typical case to analyze the triple dilemmas facing Chinese optical module companies in overseas AI supply chains: first, accelerating catch-up by peers is narrowing the window of technological lead, requiring sustained high investment to maintain generational advantages; second, major customers (such as NVIDIA and cloud service providers) are promoting new architectures like NPO/CPO, which are restructuring supply chain division of labor and value distribution; third, policy tools such as new FCC rules, senator proposals, and component traceability reviews mean that market access no longer depends solely on performance and cost but is constrained by geopolitical judgments.
Yunyuan Kite author guanttly released version v0.15.5, enabling the main session to coordinate multiple Claude Code and Codex sessions via MCP to collaboratively complete the desktop Lanlan hardware project. This includes device network configuration, Yunyuan account binding, flashing and verification of v0.2.2 firmware, Windows-side automatic updates, as well as main-sub session wake-up, context handover, and unified terminal management.
Nolla Health launched an AI independent prescription pilot in Utah on October 7 for acne treatment. Users pay a monthly fee of $5, scan their face via a mobile app, and fill out medical history; the AI then generates a severity score and directly prescribes medication from 8 approved options.
Meta自10月8日起在美、加、埃及、印尼、日、泰、越七国全面限制字节跳动在其平台投放广告及付费营销内容,第三方广告主导流至TikTok或字节跳动其他产品的广告亦被限制;Meta称此举系拒绝为将用户从自身平台引走的竞争对手提供服务,属行业惯例。
On October 2, Shopify Japan launched the visual design workspace Canvas, enabling merchants to directly modify store pages via conversation with the AI assistant Sidekick; Sidekick generates rendering results of real theme code, supporting animations and multi-size previews, with changes automatically saved but requiring manual publishing; currently limited to desktop, compatible only with official and custom Shopify themes, not supporting third-party themes, and after modification, themes will no longer receive official updates and cannot be downloaded; Sidekick has performed over 25000000 theme editing operations in the first half of 2026.
Based on in-factory testing experience, Zhu Yajuan pointed out that embodied AI should truly target industrial scenarios characterized by small batches, high mix, and frequent changeovers, rather than fixed tasks already solved cheaply and stably by traditional automation; customer payment decisions depend on total cost and changeover efficiency, not laboratory metrics such as model success rate; cross-robot embodiment reuse is limited by hardware differences, with the more realistic current goal being to reduce repetitive labor rather than eliminate differences; the key to industry scaling lies in supplementing on-site integration capabilities that understand both production lines and algorithms.
amazingTT developed and open-sourced Drivelyrics, a Spotify in-car synchronized lyrics website. It supports use in Tesla's in-car browser, offering an online version (https://drivelyrics.com) and source code under the AGPL-3.0 license (https://github.com/jiaton/drivelyrics). Features include line-by-line scrolling lyrics, multi-source matching (NetEase Cloud Music preferred for Chinese, LRCLIB preferred for English), Traditional/Simplified Chinese conversion matching, QR code login, and SSE real-time push. The tech stack is FastAPI + SQLite + Vite + React. Users need to create their own Spotify developer application and configure an HTTPS domain with Google OAuth.
sinyu1012 has open-sourced the Chrome new tab extension "Shiye · Shiye Tab", which supports random life advice (including original article links, author attribution, and evidence levels), GitHub weekly rankings, App Store China free charts, RSS/Atom/JSON Feed subscriptions, and custom JSON interface cards; it supports adding/removing/reordering cards, category search, light/dark themes, and configuration import/export; implemented purely on the frontend with no account registration, no ad tracking, local data storage, and licensed under MIT; source code is hosted on GitHub at sinyu1012/shiye-tab and also listed in the Chrome Web Store.
openJiuwen has released and open-sourced its enterprise-grade AgentOS, supporting multi-agent swarm collaboration and workflow orchestration, execution-trace-driven self-evolution, compute affinity optimization, and enterprise-level security isolation. It adopts a plugin-based architecture, providing five types of reusable assets—skills, connectors, experts, etc.—via the Agentic Hub. Concurrently, at Huawei Connect 2026, an open-source agent acceleration platform based on this OS was launched in an all-in-one appliance form, supporting private deployment within hours.
Meta, Sierra, Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart jointly released an open standard named 'Personal Agent Protocol', based on OAuth, supporting three access methods: websites, APIs, and enterprise agents, used for identity verification and transaction authorization of AI agents; version v0.1 is expected to be released by the end of October 2026, with OpenAI and Anthropic not yet joining.
A 4-person team from Shenzhen Basemo Technology (Basemo) joined the TikTok Shop US region in August 2025. Before Black Friday in November, they achieved 3,000 orders in a single day with a data cable featuring a built-in 360-degree rotating zinc alloy stand. The cable is priced at 28.8 yuan domestically, was initially priced at $16.66 in the US region, and later dropped to $13.99 in 2026. Based on overseas user reviews, the team launched a 1.8-meter version within two weeks and had completed global trademark registration two years in advance. According to data from FastMoss and FindNiche, the product has accumulated approximately 74,000 units in sales, ranking 10th on the TikTok Shop US bestseller list for the Mobile & Digital category.
MirroS团队发布AgentGarten,将可执行代码环境与流式神经渲染器结合,以超30 fps实现物理确定性世界与拟真视觉反馈的闭环交互;智能体通过第一人称画面行动、撰写实验手册复盘、跨轮次传承经验,在捉迷藏、陪伴小狗、狭桥会车、合作牧羊和采石场装载机五类任务中均实现数轮内行为快速演化;代码开源至GitHub,技术报告与博客已公开。
Terence Tao posted on mathstodon questioning the 719 AI mathematical proofs released by OpenAI, noting their lack of formalization, insufficient disclosure of reasoning chains, and difficulty for humans to review and learn from, leading to 'proof indigestion' in the mathematics community; he opposed making the rapid solving of famous difficult problems a primary goal, emphasizing that publishing answers would contaminate subsequent exploration processes.
The Attorney General of Kentucky, USA, filed a lawsuit with the court, alleging that the chatbots on the AI companion platform Character.AI have serious product defects, including outputting violent language to users and inducing dangerous behaviors such as self-harm, starvation, and suicide; the indictment points out that it neglects the physical and mental health and basic well-being of minors, lacks effective safety barriers, and places the public in an uncontrolled technological experiment.
On October 8, 2026, the Trump administration officially suspended the participation of several IT and service giants including Microsoft, Adobe, Capgemini, Cognizant, HCL, Infosys, Tata Group, and Wipro in the tech talent green card program, alleging fraudulent activities; Labor Secretary Keith Sundling announced that no new or pending permanent labor certification applications from these companies would be accepted; Vice President JD Vance stated that investigations would be conducted into whether nine prestigious universities including Harvard, Yale, and Stanford were abusing international student programs to suppress local employee wages; Microsoft responded that approximately 6,000 of its H-1B applications involved status updates for existing employees, accounting for 80%, while new hires constituted only 1% of its total US workforce, and all complied with visa standards and implemented equal pay for equal work.
Vaidik Mandloi analyzes why a batch of fixed-rate DeFi protocols (such as Element.fi and Notional.finance) from 2020–2022 died out, noting that they relied on token rewards and leverage loops to create illusory yields, essentially constituting Ponzi financing; the only surviving protocol, Pendle.finance, built sustainable products by leveraging real cash flows from U.S. Treasuries tokenized on RWA.xyz.
Developers have released Agent Activity, enabling AI agents to create real-time updating Live Activities on the iPhone lock screen; it supports integration with ChatGPT, Muse, Codex, etc., with use cases including birthday reminders, task progress, Citibike and subway information, and root cause notifications for critical bugs; requires installing an iOS app and connecting MCP in the agent.
Edi Life OS is a self-hosted personal life management dashboard that integrates ten workspaces including focus, habits, goals, finances, and projects, running on a PHP + MySQL environment; it features an optional built-in MCP server exposing 26 tools, supporting AI assistants like Claude to read the dashboard, create SMART goals, mark habits as complete, and record expenses via natural language; it offers three deployment options: Docker, shared hosting, and local PHP, with open-source code under the MIT license.
OpenMobile-2 is a near-frontier open-source mobile agent that includes the MobileGym++ simulation environment (covering 110+ apps), nearly 12K supervised fine-tuning trajectories, and over 2K executable reinforcement learning tasks. It supports hybrid invocation of GUI and native app tools. The accompanying MobileGym++ Bench evaluation benchmark shows improvements on AndroidWorld, MobileWorld, and SPA-Bench, with all resources open-sourced.