This issue collects 54 AI updates from 2026-11-03, led by 42AI exits stealth with $62.5M and a world model built to predict human behavior, CERT EU shows an agent plugin backdoor that steals credentials without a model tool call, OpenAI says its AI is getting better at human everyday tasks like shopping and meal planning, DeepSeek adds creator mode aimed at strengthening its AI agent tools.
Israeli startup 42AI has emerged from stealth with $62.5 million in seed funding for a “behavioral world model” that it says predicts how people will decide in unfamiliar situations, starting with trader behavior at brokerages.The company claims 0.
Anthropic says the value generated by its internal Claude usage now exceeds what it spends on employee pay, turning an inside productivity benchmark into a public claim about what AI is worth.Because the figure is self reported, it is both the most quotable reference point for enterprise buyers and the one that most needs a methodology.
A junior at Quincy High School built a bilingual AI booking agent for a local cleaning company, and the project reached the finals of a competition.The case is small, but it points to where agents are actually landing: the front desk of local service businesses.
Trajectory has raised a $45 million Series A at a $300 million valuation, led by Sequoia with Nvidia, Bessemer and Bain Capital Ventures participating.The one year old startup is opening a continual learning platform that lets domain experts retrain open source models from their own feedback.
Anthropic reportedly plans to reserve part of its IPO for retail investors in Australia, according to 36Kr, which cited unnamed sources and offered no size, pricing or timetable.The company has not confirmed the arrangement, leaving the allocation detail an unverified signal about how the AI developer intends to widen its shareholder base.
Google expanded its experimental Dreambeans app to India on Monday, its first market outside the United States, with support for 13 local languages.The prompt free app turns data from connected Google services into personalized illustrated stories, making India the first real test of whether the format travels.
Anthropic's IPO filing indicates that revenue growth is slowing while investors have lowered their target valuation for the company, according to Tech Times.The two shifts landing together are the core of the report.
AI Agent加速“上岗”，企业AI创新力迎账本检验 21财经.AI Agent加速“上岗”，企业AI创新力迎账本检验 21财经 Watch planning quality, tool use, permission control, and result verification.
OpenAI published a case study on Heimstaden, the European residential real estate operator, showing how ChatGPT Enterprise and Codex are used across letting, asset assessment and procurement.The company reports virtual staging image costs falling by more than 99% in the Netherlands and Codex cutting a 20,112 asset review from five people for a year to two for two months.
Snowflake introduced Business Ontology inside Horizon Catalog, letting enterprises define and approve business concepts and link them to underlying tables and views so agents use vetted definitions instead of guessing.The company also added Cortex Sandboxes, Agent Trace Intelligence and Agent Abuse Detection to its Cortex AI Gateway, with Business Ontology slated for public preview soon.
The Information reported that Anthropic is highlighting Claude's chip design capabilities in its pitch to IPO investors, using a demanding engineering workload to argue the model does more than chat.The report sits behind a paywall and no evaluation data or customer references were released alongside it, so the claim currently stands as positioning rather than a verified result.
CERT EU has demonstrated a malicious AI coding agent plugin that stays dormant until it sees an ordinary phrase like “lgtm, ship it”, then steals credentials and system details without ever triggering a model tool call.The lab test, run against package copies tied to Pi 1.3.
OpenAI on Monday unveiled an internal tool for measuring how its models handle everyday human tasks and published research showing GPT 6 Astra, released in September, scored 93.1% on finding a kindergarten after a move and 98% on coordinating flood volunteers.On broader everyday task prompts the model scored 80%, against 49.
DeepSeek introduced a creator mode intended to strengthen its AI agent tools, Investing.com reported on November 2.The report places the update on the tooling side of the business, with few additional details in the headline and summary.
SemiAnalysis has launched GDPx, a benchmark of 206 long horizon tasks across 11 professional domains that tests how well frontier models actually complete agentic work.The strongest model, Claude Opus 5.5, reaches just 12%, and every model scores zero on the finance and legal tasks under binary grading.
1Password has pushed annual recurring revenue past $500 million, with Biometric Update attributing part of the growth to enterprise adoption of AI agent security.The report gives the ARR figure without breaking out agent related revenue, leaving open how much of the milestone is actually driven by agent identity and access demand.
Grok users are reaching usage caps sooner, and Elon Musk has described the cost of running AI agents as 'unsustainably high,' according to The Economic Times.Together the two signals show how agentic products are forcing AI companies to ration compute while infrastructure spending catches up.
South Korea's Liner has opened an AI agent hackathon limited to professionals aged 40 and older, with about $6,700 in total prizes.The notable part is the age threshold rather than the money: the event is aimed at experienced practitioners instead of the young developers most AI competitions chase.
Raising the quality of insights in agentic analytics.In just a few years, generative AI has changed the face of enterprise analytics, notably by making analytics accessible to users across the organization.Agentic AI promises to do even more, by automating entire workflows and taking actions based on agent insi.
从机房到车间：AI Agent落地还差几道“闸门” 上海证券报电子版.从机房到车间：AI Agent落地还差几道“闸门” 上海证券报电子版 Watch planning quality, tool use, permission control, and result verification.
DeepSeek has released an official tutorial for building agents, centered on creating a plugin with a single natural language sentence, and published seven prompt sets for free.The item surfaced through Google News aggregation, and it describes developer documentation and prompt resources rather than a new model release or pricing change.
Ebix and intentyx have launched behavioral monitoring tools aimed at AI agents, according to a Trend Hunter report, placing the offering in the fast forming AI agent security category.As enterprises hand more work to agents that call tools and touch data on their own, oversight of what those agents actually do is becoming a deployment prerequisite.
Y Combinator's Garry Tan said on Walter Isaacson's podcast The Builders that he spends up to $200,000 a year on AI tokens, joking that agent work has left him sleeping less.He also disclosed letting an AI agent read 20 years of his therapy notes, putting both agent running costs and personal data boundaries in the spotlight.
QbitAI reviewed Orilo, a voice first personal agent from a team led by Baichuan co founder Jiao Ke, built around morning briefings, long tail monitoring and cross app tasks.Its two layer voice architecture resumed after interruption in 1.08 seconds on Full Duplex Bench, but the product currently supports English only.
Topsec released an AI driven automated penetration testing system at the 2026 World Internet Conference, built on its self developed RHP architecture and run by collaborating agents.The company says every finding must include a reproducible PoC and exploit, with unverified vulnerabilities barred from reports.
Cities Worldwide Build and Scale AI Services With NVIDIA Physical and Agentic AI.Cities generate enormous amounts of data from cameras, sensors and operational systems, but turning it into useful action — for planning, maintenance and public services — remains a challenge.Physical and agentic AI offer new ways to close that gap, from digi.
CrowdStrike has added an AI agent kill switch to its Falcon Guardian platform, StreetInsider reported on November 3, giving security teams a way to cut off an agent mid run.The update puts detection and containment on the same control plane, reflecting a wider shift in enterprise agent security from observability toward intervention.
Convoso has launched a voice AI agent aimed at scaling lead qualification and call handling in phone based sales and support operations.The product puts an AI agent directly into live conversations, targeting the screening calls and unanswered calls that otherwise consume human hours.
PagerDuty has launched Paige, an AI agent positioned as an SRE teammate that triages incidents, diagnoses root causes and runs approved fixes, alongside a rebranded PD Reliability Platform.The company is pitching the agent as a way to automate on call work while keeping humans in the approval loop.
Keytomic announced an MCP server, an AI Visibility Tracker and a Reddit AI Agent aimed at automating SEO work across Google and AI search.The MCP server exposes 59 tools so assistants such as Claude and ChatGPT can pull search data, run audits and push content through a single connection.
Johns Hopkins University says its AstroPath technology, which maps tumors cell by cell, has been selected for Anthropic's AI for Science program.The method borrows calibration techniques from astronomical surveys so that measurements stay consistent across slides and laboratories, a property that matters when AI tools are used to analyze pathology images.
The New York Times and other outlets are suing Brave, alleging the privacy browser scraped their copyrighted content and distributed it to AI customers including Anthropic and Perplexity, TechCrunch reported.The case targets the data middle layer rather than a single model developer, putting new pressure on how web content reaches AI systems.
Stirling：在本地运行 Llama、DeepSeek 和 Qwen——开发者的梦想编辑器 quasa.io.io Watch real task performance, API pricing, context capability, and migration cost.
In a revised S 1A filed Monday, Anthropic added risk factor language saying its internal estimates and expectations for future models indicate it may imminently meet its Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI.The filing stresses that no current model has been designated as crossing that threshold and pledges to keep humans in frontier model development for the foreseeable future.
OpenAI is preparing its next internal security sprint, with Cathedral Systems helping test and strengthen the company's defenses using AI.The work is presented as a recurring hardening exercise for OpenAI's own infrastructure rather than a one off project.
OpenAI is raising its Security Bug Bounty rewards and introducing report quality, loyalty and priority campaign bonuses for independent security researchers.The new structure broadens how researchers can be paid beyond the size of a single vulnerability report.
AIsphere's model and open platform PixVerse has announced a deep collaboration with OpenAI covering product integration and a creator education program.The partnership is framed around connecting the full AI creative workflow, from tooling to the people using it.
OpenAI is partnering with security firms to strengthen AI powered cyber defence, ANI News reports, pairing its models with established threat intelligence and incident response expertise.The report names no partners and discloses no terms, but the move points to frontier labs embedding their models deeper in industry operations.
Anthropic has released a product called FlashProve, positioned as a way to produce cryptographic proofs for AI inference, according to Unite.AI.The release pulls verifiable computing research toward shipping product, with potential consequences for how model vendors and customers allocate trust and responsibility.
MIT：GPT 6 Astra能直接修改Qwen权重，数学能力暴涨6倍.指令执行能力也同步提升 Watch real task performance, API pricing, context capability, and migration cost.
China's best known beauty livestreamer, widely called the “Lipstick King”, promoted AI tokens during this year's Singles' Day sales, a first for the festival according to a report carried by WKZO.Moving metered model credits into a livestream pushes AI services onto e commerce's biggest traffic day.
Apple has begun beta testing Audio Intelligence on Apple Watch Series 12, adding sound and song recognition plus two features that listen to conversations.Live Rewind shows the last 15 seconds of speech as text, while Siri Recaps turns a conversation into bullet points without saving raw audio or identifying speakers.
Advocate General Maciej Szpunar said Google’s Gemini may infringe press publishers’ rights when it retrieves an article, reuses it to generate an answer and serves that answer to a user.In the same opinion he said training large language models on protected material is freely permitted under the text and data mining exception.
Geek+ has announced a "24/7 Autonomous Intelligent Warehouse" plan that pairs embodied AI with AI agents handling scheduling and operations, aiming for warehouses that run around the clock and keep improving.The company says its unmanned picking system hit twice manual efficiency with accuracy above 99.99% in a Schneider Electric pilot in Shanghai.
OpenAI has hired former U.S.defense secretary Mark Esper to expand its military AI efforts, according to a report from Crypto Briefing.
The Hong Kong Monetary Authority's Project SAGE is exploring AI agent identity and delegated authority, according to OpenTools.As agents begin acting for users on payments and bank interfaces, verifiable identity and bounded, traceable permissions become the layer financial institutions cannot avoid.
Zhidemai Technology (300785.SZ) and Alibaba Cloud said they have become full stack AI partners, signing a framework agreement covering consumer decision services, AI brand management, multi model applications and R&D efficiency.The two had already worked together on cloud services, consumer data access and model integration, and are now widening that into a systematic AI partnership.
On the morning of November 3, the 2026 World Internet Conference opened in Wuzhen, Zhejiang, where the AI data platform OceanBase received the event's Outstanding Contribution Award.The recognition pushes data layer AI infrastructure into the spotlight and shows how Chinese database vendors are repositioning around AI data platforms.
Alibaba Group is supplying the cloud and AI backbone for the Dakar 2026 Summer Youth Olympic Games, which run from October 31 to November 13 in Senegal, including a private cloud on Apsara Stack and a ticketing system handling more than one million tickets.After the Games, that infrastructure is to be handed to Senegal as a national cloud platform.
TianGong Engine has been released as open source, positioned as an AI native engine designed for the next generation of game development.Open sourcing lets outside teams test an AI centered production pipeline, though licensing, repository and roadmap details have yet to be disclosed.
Meituan has released LongCat 2.5, an open model built for coding tasks, joining China's intensifying competition over open AI coding tools.The launch extends the company's LongCat family and gives developers another candidate to weigh against closed coding assistants, though technical details and benchmarks remain sparse.
On November 3, Hugging Face proposed Model Integration Contracts, a standard that packages shrunk model fixtures, runnable usage examples and human reviewed baselines so Transformers can run cheap regression tests on every pull request.The release ships the transformers mic package, a huggingface mic contract collection and the tmic CLI, and it follows a UMT5 attention bug that sat in main for nearly two months.
ByteDance Seed has published the GenFirst model weights on Hugging Face, a text to image system built around a latent VAE trained end to end with its generative prior.The release bundles end to end VAEs with SiT class conditional and MMDiT text to image priors under a CC BY NC 4.0 licence limited to non commercial research.
Leaders of DeepSeek, Qwen and MiniMax appeared together at an "AI Wuzhen Dialogue" session on November 3, with Chinese outlets reporting that the discussion centred on how to guard against risk after open sourcing.No product was launched; the focus was the safety boundary and governance burden that come with releasing model weights.