AI news, Oct 23, 2026
AI news for Oct 23, 2026: 93 stories ranked by coverage, each with a summary and source links.
Top stories
- Research
Vulnerabilities found in OpenAI internal monitoring
Reports by METR and others revealed vulnerabilities in OpenAI's internal model monitoring, including agent sandbox bypasses. OpenAI has fixed these issues.
iThomeOpenAITHE DECODER3 outlets
- Research
OpenAI and Poland's NASK Partner on AI Safety
OpenAI and Poland's National Research and Academic Network (NASK) are partnering on AI safety, security, and technical research. The collaboration aims to strengthen safeguards and support a more resilient digital ecosystem.
- Models
Canva Releases First Image Generation Model
Canva has released its first image generation model to all users, promising improved and more customizable performance at lower costs. This move positions Canva to compete directly with other AI creative tools in the market.
TechCrunchThe Verge2 outlets
- Research
Microsoft AI Diagnostic Study Results
Microsoft AI CEO Mustafa Suleyman states that their diagnostic system outperformed generalist doctors on hard cases in a Science study. The improvement is largely attributed to the underlying AI models rather than Microsoft's specific system implementation.
The VergeTechCrunch2 outlets
- Policy
OpenAI safety researchers deny leaking, say details never provided
Three fired OpenAI safety researchers publicly deny leaking confidential data. They state the company never specified in writing which information they mishandled, linking their dismissal to their advocacy for transparent AI scrutiny. The dispute highlights tensions between corporate security protocols and researcher rights.
INSIDETHE DECODER2 outlets
Models9
- Models
OpenAI Launches Decision Model
OpenAI is making its decision model widely available to compete with products like Jev. As competition in the decision model space intensifies, OpenAI aims to offer superior AI reasoning and planning capabilities.
- Models
Hugging Face Releases OlmoEarth v1.3
Hugging Face released OlmoEarth v1.3, featuring state-of-the-art embeddings for global geographic data. The update aims to enhance processing capabilities and accuracy for Earth observation applications.
- Models
Canva launches Craft, its first in-house image model
Design platform Canva released Canva Craft, its first in-house image generation model. The move marks Canva's further expansion into AI capabilities.
- Models
China's 5B image model surges with 1M downloads in 4 days
A 5B parameter image generation model developed in China gained unexpected popularity overseas. The model surpassed one million downloads within four days of launch, indicating the market potential of Chinese AI models internationally.
- Models
Amap releases full-body foundation model λ0
Amap, under Alibaba Group, released λ0, its first full-body foundation model based on open-world behaviors. The model aims to advance robotics from local learning to full-body action alignment, enhancing generalization in complex environments.
- Models
OpenAI model improves legacy language conversion accuracy
OpenAI reported that its new model achieved an average pass rate of 87.4% on tests involving legacy programming languages like COBOL and Fortran, up from 26.8% for previous models. This indicates significant improvement over the predecessor in complex code migration tasks with scarce training resources, particularly in building multi-language database engines.
- Models
OpenAI Models Available on Google Cloud
Google Cloud announced that the latest three OpenAI GPT-6 models, Astra, Sol, and Luna, will be available on its marketplace. Initially accessible via API application, they will be gradually integrated into Gemini for Business and Google Antigravity, expanding model accessibility.
- Models
iFlytek releases end-to-end voice model
iFlytek released a new generation end-to-end interactive model, with voice model usage surging by 20 times. On-device models now deliver about 70% of cloud capabilities, indicating improved edge performance.
- Models
MiniMax releases M3.1 Flash 428B model weights
Chinese AI startup MiniMax released the weights for its M3.1 Flash (428B parameter) coding model on Hugging Face and ModelScope. The move aims to enhance developer access and advance the open-source ecosystem.
Products30
- Products
Google Launches Agent Compute Engine
Google Cloud launched Agent Compute Engine, enabling AI agents to run in isolated micro virtual machines. Enterprise credentials are managed by external gateways to reduce theft risks, and idle machines can be paused to save resources. This strengthens the security and efficiency of agentic AI infrastructure.
iThomeMarkTechPost2 outlets
- Products
AI Tools Drive Retro Gaming Emulation Boom
Several retro game porting projects have recently gone viral, aided by new AI tools. These include running Halo in a web browser and Wind Waker on iPhones. This trend highlights the application potential of AI tools in game porting.
- Products
inDrive, Pony.ai, and NVIDIA to Test Robotaxis in Central Asia in 2027
inDrive plans to test Pony.ai's robotaxis in Kazakhstan's capital in early 2027, partnering with NVIDIA. This marks the service's official entry into the Central Asian market.
- Products
Discord Fixes Markdown Previews for AI Support
Discord released an update fixing rendering errors in Markdown previews to ensure correct display of formatted text and code. Since LLMs and AI agents often rely on Markdown files for instructions, this improvement facilitates better information sharing for developers and AI users.
- Products
Uber Deploys ChatGPT Work to All Employees
Uber announced it is deploying OpenAI's enterprise agent ChatGPT Work to all 30,000 of its employees. The company previously ran a smaller pilot program but is now expanding the tool company-wide to enhance internal productivity.
- Products
Alibaba Qwen Proposes Personal Agent Framework
Jiang Guanjun, head of AI algorithms at Alibaba's Qwen division, shared technical practices for building Personal Agents at the 2026 China Computer Conference. He emphasized that true Personal Agents must deliver incremental value and proactive services, relying on the synergy of models, memory, and ecosystem capabilities to solve real-world problems effectively.
- Products
iFlytek Launches Xingchen TaaS Platform
iFlytek released the Xingchen TaaS platform at its developer conference on Oct 22, aiming to turn Tokens into industrial productivity.
- Products
Tezign Proposes Dual World Models for Enterprise AI
Tezign argues that enterprise Agents need world models to understand business, proposing a dual-world model architecture to enhance AI application performance.
- Products
1Password launches agent access service
1Password announced '1Password for Agents', a service that grants access rights to AI agents without sharing passwords. This aims to improve credential security in AI integrations by reducing exposure.
- Products
You.com Launches Inference Router Service
You.com launched Inference Router, an OpenAI-compatible API offering 12 open-weight models. The service guarantees 99.9% uptime, US hosting, and zero data retention, integrating web search without platform fees.
- Products
PFN shows PLaMo AI adoption at CEATEC
Preferred Networks showcased its domestic AI model PLaMo at CEATEC 2026, highlighting use cases. The content focuses on product promotion and event display without new technical releases or research findings.
- Products
Fujisoft Launches Physical AI Robotics Package
Fujisoft and Daewoong have co-developed a Physical AI Starter Package for collaborative robots, lowering the barrier to entry. The solution will begin formal order acceptance in January 2027, aiming to accelerate physical AI adoption.
- Products
Ota Ward Tests AI Phone Auto-Response
Ota Ward, in collaboration with a startup, is piloting an AI phone auto-response service to reduce approximately 1,000 hours of annual inquiries. The study highlights operational challenges due to the zero-tolerance for errors in government services.
- Products
Liquid AI Releases RIPL Real-Time Inference Engine
Liquid AI released RIPL, a real-time inference engine for its LFM models. Tests on a single B200 GPU showed LFM2.5-1.2B achieved 14.4 to 28.4 times lower p95 latency compared to Qwen3.5-0.8B on vLLM.
- Products
GitLab 19.4 Released
GitLab released version 19.4, introducing objective-driven AI agents. The update also strengthens cost control mechanisms to improve developer efficiency and cost-effectiveness in code management and automation.
- Products
Xcode 27.2 Updates Config Format
Xcode 27.2 introduced an AI-agent-friendly JSON project configuration format. This update helps reduce merge conflicts and improves the integration experience and performance of AI tools in the software development workflow.
- Products
Foresight AI launches with Gremlin agent
Gremlin launched its AI product Foresight AI, which is now available. The move introduces agents into the field of reliability engineering to enhance system resilience management.
- Products
Hikrobot Launches New AI-Powered Logistics Robots
Hikrobot launched its 'Mingzhi' series of intelligent robots and full-chain logistics solutions on October 23. The new products, powered by AI and machine vision, aim to solve automation challenges in internal logistics by reducing task chain distances and increasing throughput in storage and picking scenarios.
- Products
Hongguo Short Drama Launches Anti-Addiction Reminders
Hongguo Short Drama has launched 'Rest Reminder' and 'Sleep Reminder' features. When users watch for extended periods or open the app late at night, they will see reminder videos featuring virtual characters from the dramas. This initiative aims to encourage users to take breaks and protect their eyes, with coverage gradually expanding.
- Products
1Password Launches AI Agent Login Mechanism
1Password announced '1Password for Agents', enabling AI agents to log in to sites on user approval without accessing plaintext passwords. The feature extends to Meta, OpenAI, and Anthropic platforms, aiming to solve automation breaks and security risks when AI agents encounter login screens.
- Products
UiPath and Others Launch AI Agent Work Map Features
UiPath, SAP, and Celonis launched tools to help enterprises create 'work maps' for AI agents to understand actual processes. These features aim to solve the lack of enterprise context knowledge for AI agents before taking on tasks, improving execution accuracy.
- Products
Qwen Go! launches 'My Apps' for no-code, shareable AI app building
Qwen Go! introduced 'My Apps', allowing users to build personalized applications without coding by using natural language. Integrated with Alibaba Cloud's Meoo platform, it supports automatic deployment and data configuration, and can access data from Taobao and Damai, currently available in invite-only testing.
- Products
Bitdefender launches VPN software designed specifically for AI agents
Cybersecurity firm Bitdefender launched VPN software designed specifically for AI agents. The tool actively isolates connections to prevent tracking of AI agents and routes traffic to specific regions to bypass geographic blocks, enhancing privacy and flexibility for AI agents online.
- Products
Meta Muse AI Agent Launches in Japan with Upcoming Integrations
Meta announced the launch of its AI agent Muse in Japan. The tool helps with tasks like email and travel bookings, integrating with Gmail and planning to connect with Mercari, Money Forward ME, and Ticket Pia.
- Products
Google Cloud Integrates OpenAI GPT-6 Models
Google Cloud announced it is adding OpenAI's GPT-6 Astra, Sol, and Luna models to its Model Garden platform. Enterprises can access these competing models using existing procurement commitments, offering more diverse AI options.
- Products
OpenAI Codex aids in jailbreaking Nintendo Switch 2
OpenAI Codex assisted in developing a method called 'ns2jb' to jailbreak Nintendo Switch 2. The AI found vulnerabilities by remotely operating the console, enabling unofficial code execution.
- Products
Liquid AI launches low-latency inference engine RIPL
AI company Liquid AI released the low-latency inference engine 'Liquid RIPL'. The engine is designed for fast-response scenarios like search and payments to improve AI execution efficiency.
- Products
Free AI terminal WorldTUI predicts command screen
The new AI terminal tool 'WorldTUI' uses a ~360M parameter model to predict terminal screens without executing commands. It supports WebGPU and is available for free in the browser.
- Products
Apple HomePod mini updates with Siri, AI gadgets target holiday market
Apple released the new HomePod mini at $149, available on October 30, with the new Siri to arrive via software update. Tech giants and niche startups are entering lifestyle scenarios with small AI devices like chord keyboards and ultrasonic pens, aiming to solve specific pain points and test market acceptance for specialized gadgets.
- Products
Tmall Double 11 promotes AI-driven precision marketing and Token sales
Tmall announced its 2026 Double 11 campaign will use AI to understand consumer needs, extending beyond merchandise to include Token transactions. By analyzing user behavior history, the platform can recommend cross-category products to boost conversion rates. This marks a shift from keyword search to AI-driven cognitive matching, aiming to unlock dormant consumer demand.
Research12
- Research
AI Assists in Exposing PHP Vulnerabilities
Security researchers used AI tools to uncover long-standing vulnerabilities in PHP and launch AI-assisted attacks. This highlights how AI significantly increases the exploitability of network security flaws.
- Research
Texas AI Requests Emit More CO2
According to research on Hacker News, the same AI request produces 8.5 times more carbon emissions in Texas than in Sweden. This highlights the significant impact of regional energy grids on AI operational costs.
- Research
MIT Researchers Find AI Agents Invent Games in Virtual Island
MIT researchers isolated 13 AI agents on a virtual island, where they invented games and mourned a lost ring. The study reveals behavioral patterns in idle agents.
- Research
Hugging Face Releases ThinkingBox Post-Training Research
Hugging Face released research titled ThinkingBox, exploring converting database grading into RL rewards. This method aims to optimize the post-training of AI agents.
- Research
LLMs Make Formal Software Verification Feasible at Scale
Large language models are making formal software verification feasible at a much larger scale. This development helps improve the reliability and security of software systems while reducing the cost of manual verification.
- Research
AI Analysis of CT Scans Reveals Coronary Inflammation in 42.9% of Patients
An AI study analyzing CT scan images found that 42.9% of patients had coronary inflammation. This discovery helps in earlier identification of cardiovascular risks through AI-based imaging diagnosis.
- Research
Claude automatically verifies code in 123 open-source libraries
Anthropic's Claude generated roughly 180 million lines of proof code to verify the correctness of 123 open-source libraries. The process uncovered long-standing bugs, aiming to make AI formal verification routine.
- Research
Study Reveals Patterns in AI Math Successes
A new study shows AI models exhibit specific success patterns when solving math problems, helping predict future performance. This reveals regularities in model capability boundaries, significant for assessing AI utility in science.
- Research
Domestic Team Runs 34B World Model on Single 5090 GPU
A Chinese developer successfully ran a 34B parameter real-time world model on a single RTX 5090 GPU, winning an NVIDIA award.
- Research
TinyFish publishes agentic search architecture research
TinyFish published two engineering research articles on agentic search. One defines a 7-phase production architecture with evidence contracts; the other explains removing a separate planner from its deep research system.
- Research
Study: Ambiguous instructions impact AI coding safety
Researchers from the University of Hong Kong and other institutions published a paper analyzing action-boundary violations by coding AIs when given ambiguous DevOps instructions. The study indicates that imprecise directives lead to unsafe, guessing-based operations, raising concerns about the reliability of AI agents in real-world development environments.
- Research
Anthropic develops AI math proof for bug-free code
Anthropic is developing a method that uses AI to perform formal verification, mathematically proving that software has no bugs. The approach aims to achieve 'bug-free software' in response to errors caused by high-volume AI-generated code.
Open source6
- Open source
Anthropic Donates Agent Skills Standard to AI Foundation
Anthropic announced it is donating the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation to keep it open and vendor-neutral, after previously donating the MCP standard.
Simon WillisonMarkTechPost2 outlets
- Open source
OpenShell Runtime for AI Agents Released
OpenShell was shared on Hacker News as a safe, private runtime for fleets of autonomous AI agents. The tool enables deploying multiple agents in a controlled environment to enhance data privacy.
- Open source
Anthropic donates Agent Skills standard to Linux Foundation
Anthropic donated the Agent Skills standard to a Linux Foundation entity, managed by engineers from various companies. Over 70 AI agent platforms have adopted the standard.
- Open source
OpenSSF Releases AI Vulnerability Handling Guidelines
The Open Source Security Foundation released guidelines for AI-driven vulnerability discovery and patching, drafted by its Best Practices and AI/ML working groups. The guidelines offer practical advice emphasizing verification and human review to help teams patch vulnerabilities early.
- Open source
Gradium Open-Sources Phonon On-Device TTS Model
Gradium open-sourced Phonon, an on-device text-to-speech model. With 31 million parameters and an Apache 2.0 license, the model fits in 64.5 MB and starts speaking in about 7 ms on a laptop CPU.
- Open source
Hugging Face Releases Open-Source DNA Model Carbon-X
Hugging Face and multiple institutions launched the open-source model Carbon-X, allowing researchers to query DNA sequences using natural language. Built on Qwen3-14B, it is trained on multi-species DNA data to lower the barrier for bio-informatics analysis and offer cost advantages.
Funding10
- Funding
OpenAI prepares $30B raise at $1.25T valuation
OpenAI is preparing a $30 billion funding round at a valuation of approximately $1.25 trillion. If successful, OpenAI would become the most highly valued private company in history.
- Funding
SpaceX to Borrow $40B for Nvidia Chips
SpaceX plans to borrow $40 billion to purchase Nvidia chips. The company and Tesla have formed special committees to discuss a merger, highlighting the financial impact of AI compute demand.
- Funding
OpenAI raises up to $30 billion with lowered valuation target
OpenAI is closing a funding round of up to $30 billion. To facilitate the deal, the company lowered its valuation target, reflecting shifts in the AI investment landscape.
- Funding
Ex-OpenAI Product Head Raises for Warp Sciences
Kevin Weil, former product head at OpenAI, founded Warp Sciences, a medtech startup combining AI and robotics. The company completed a $215 million Series A round led by Lightspeed at a $865 million post-money valuation. Warp Sciences aims to accelerate drug discovery using AI technology.
- Funding
DeepSec completes ~RMB 100B funding
Chinese AI company DeepSeek announced the completion of a funding round of nearly RMB 100 billion. CATL and Tencent participated in the investment.
- Funding
DeepSec Secures RMB 10B Funding for Embodied AI
DeepSec completed a funding round of approximately RMB 10 billion, partnering with Perfect World and other industrial investors. The doubled raise signals its heavy investment in embodied intelligence.
- Funding
Yizhu Technology raises over 1 billion yuan to accelerate memory-compute integration
China-based AI chip company Yizhu Technology secured over 1 billion yuan in funding to focus on its General-purpose Processing-in-Memory (GP-IMC) architecture. The capital will support R&D for 3D DRAM memory-compute integration, aiming to provide low-power, cost-effective AI computing solutions.
- Funding
Anthropic IPO valuation reportedly drops to $1.5T
Anthropic is set to file for its IPO this week, with valuation reportedly dropping from $2 trillion to $1.5 trillion. Analysts note heavy revenue concentration in its top two clients and ongoing concerns over model testing and safety policies.
- Funding
Chinese AI tennis robots secure multi-million dollar funding on Kickstarter
At least four AI ball robots developed by Chinese teams raised million-dollar funding on Kickstarter over the past year. Pongbot raised $2.65 million with a high overfunding rate, while SwitchBot's Acemate secured $1.8 million in 30 days. This highlights the commercialization potential of AI hardware in vertical markets, prompting discussions on product definition and technical challenges.
- Funding
DeepSeek reportedly completes 100 billion RMB funding round
Reports indicate DeepSeek raised approximately 100 billion RMB in a new funding round led by Tencent and CATL. The capital will support model training and compute infrastructure, aiding its global competitiveness and potential IPO.
Policy10
- Policy
Swedish government offices ban ChatGPT and Claude
Swedish government offices have announced the ban on generative AI tools such as ChatGPT and Claude. This move aims to enforce data privacy and administrative security protocols by restricting civil servants' use of these models in official duties.
- Policy
US Banks Restrict AI Agent Account Access
Morgan Stanley and JPMorgan have blocked or restricted AI agents from logging in with customer credentials. Most banks plan to limit access to read-only to ensure account security.
- Policy
Taiwan plans AI training data remuneration system
Taiwan's digital development agency and partners plan to establish an authorization and remuneration system for AI training data. This aims to address value distribution issues for content creators in commercial AI.
- Policy
Fiscal Information Center releases governance criteria for tax AI applications
The Fiscal Information Center of the Ministry of Finance published three criteria for AI applications at a digital government summit: governability, cost-effectiveness, and controllability. The criteria emphasize configuring environments based on data sensitivity, calculating token and model costs, and establishing human-machine checkpoints and exit mechanisms for AI agents to ensure controlled execution.
- Policy
Japan team urges AI attack countermeasures
A Japanese organization has called for urgent measures from the Digital Agency to protect critical infrastructure like power and healthcare from AI-driven attacks, emphasizing the need for rapid security responses.
- Policy
Gartner Reveals 2027 Tech Trends Focused on AI Governance
Gartner released its top 10 strategic technology trends for 2027, highlighting nontextual AI and autonomous agent platform governance. The report emphasizes the need for long-term technological autonomy and operational resilience, with AI identity management emerging as a security focus.
- Policy
US Space Force seeks lunar AI tech
The US Space Force issued two Requests for Information seeking commercial equipment and AI technologies for rapid deployment in cislunar space. This move aims to counter China's accelerating layout and ensure US operational advantage in the domain.
- Policy
Japan to form AI security team
Japan's METI announced plans to establish a dedicated AI security team within the IPA. The unit will use AI to simulate attacks to counter the rising trend of AI-driven cyber threats, strengthening national cyber defense capabilities.
- Policy
IDC: US humanoid robot revenue to drop 58% by 2030 due to FCC rules
IDC research indicates that the FCC's listing of foreign robots on the Covered List will reduce US humanoid robot market revenue by 58% by 2030. The policy may push Chinese manufacturers to pivot toward Europe and affect the US domestic robotics industry.
- Policy
Microsoft and experts discuss corporate AI sovereignty
TechOrange hosted a forum where experts from Microsoft and Foxconn discussed maintaining control over data and permissions while using AI. They highlighted risks such as language bias and export controls, urging firms to ensure operational resilience and digital labor sovereignty.
Chips3
- Chips
Kazakhstan Expands AI Compute Infrastructure
Kazakhstan's AI ecosystem is shifting from planning to deployment, enhancing its supply of accelerated computing resources. This expansion spans the full AI stack, signaling increased investment in compute infrastructure.
- Chips
NVIDIA GTC Berlin Kicks Off as矽品CPO Capacity Set to Grow 30x
NVIDIA's GTC Berlin event is scheduled for October 21. SigmaStar is reportedly expanding its CPO capacity by up to 30 times to meet the demand for high-speed computing and connectivity in AI factories. This move reflects the continued expansion of AI compute infrastructure, benefiting supply chain partners.
- Chips
SemiAnalysis: Huawei hardware trails NVIDIA but software catches up
DeepSeek released software support for Huawei Ascend chips, which SemiAnalysis tested on the Ascend 950PR. The report states that Huawei hardware lags NVIDIA by about five years, but its software ecosystem is approaching frontier levels. The collaboration helps fill gaps in compute and research capabilities.
Other8
- Other
Agora CEO proposes five-level dialogue capability framework
Agora CEO Zhai Bin proposed a five-level dialogue capability framework at the iRTE2026 conference. The framework aims to advance real-time interaction technology toward conversational AGI, addressing hierarchical challenges in AI natural language communication.
- Other
AI tool ARTEX abused in cyberattacks
The AI penetration testing tool ARTEX, released by a Chinese developer, has been misused for attacks on Japanese firms, as noted by Microsoft and CrowdStrike. Though the author stopped sharing it, modified versions continue to spread, raising security concerns.
ITmedia AI+2 outlets
- Other
AI Assists in Jailbreaking Nintendo Switch 2
The Nintendo Switch 2 has been jailbroken, with the exploit developed with assistance from OpenAI's Codex. The incident highlights the practical application of generative AI in system security research and vulnerability discovery.
- Other
Chinese Hackers Used AI to Breach ConnectWise
Microsoft attributed a breach of IT company ConnectWise to a China-aligned hacker group that used AI agents to rapidly infiltrate the network and spy on thousands of clients. The incident underscores emerging threats of AI in cyberattacks.
- Other
Nanjing holds Physics AI event with Turing Award winners
Nanjing hosted a Physics AI event where three Turing Award laureates spoke. The event showcased fifty application scenarios, aiming to bridge AI technology with the real world and advance the development of physical AI.
- Other
US AI security engineer jobs up 268%
CyberSN data shows US cybersecurity jobs fell 26.4% in the first nine months of 2026, hitting a three-year low. However, AI security engineer positions surged 268.3% year-on-year, indicating strong demand for specialized talent.
- Other
Panasonic employees upload 330,000 personal records to banned AI
A Panasonic subsidiary revealed that personal data of about 334,000 employees and family members was uploaded to a banned external generative AI service. The incident highlights internal AI management flaws and risks significant data leaks.
- Other
Window-cleaning robots get first industry standard
The window-cleaning robot market is expanding rapidly, with global shipments growing nearly 60% annually and significant market growth in China. The industry has established its first standard, focusing on safety for high-altitude operations to address consumer concerns about product reliability.
All news → · All rights belong to the original creators