This issue collects 46 AI updates from 2026-11-05, led by Atlas Cloud Adds ByteDance's Seedance 2.5 as Video Model Distribution Battle Widens, Tencent Ships Upgraded Hy4 Model With Visual Understanding, Anthropic Open Sources FlashProve to Verify Which Model Answered, CrowdStrike buys Fiddler and Dreadnode to test, monitor and stop AI agents.
Atlas Cloud said it has integrated ByteDance's Seedance 2.5 AI video model into its platform, letting developers and content teams call the model through a single interface layer.The announcement came as a press release and did not disclose pricing, availability regions or a launch timetable.
A report says TeraWulf's $19 billion lease with Anthropic has obtained a backstop from Google, tying a capacity landlord, a model developer and a hyperscale platform into one deal.The headline level disclosure does not yet spell out the lease terms or the exact scope of Google's guarantee.
Tencent has released an upgraded Hy4 model that adds visual understanding, according to Caixin Global.The report stays at the announcement level, leaving parameter counts, context length, release format and pricing undisclosed.
TechCrunch's hands on review finds Apple's AI powered Siri Recaps surprisingly useful after a single day of use, while leaving the author uneasy about privacy.The piece credits the feature with making information review noticeably effortless, but says it still has clear room for improvement.
Ripple 65B 对比 DeepSeek V4.1 Pro：其中一个并不存在 OrcaRouter.1 Pro：其中一个并不存在 OrcaRouter Watch real task performance, API pricing, context capability, and migration cost.
Nothing CEO Carl Pei said on X that the company's first 'AI native product' ships in December, describing it as a productivity tool with no new screen that could not have been an app.He also pushed back on the idea that it will be a pendant, a pin or a pair of AI glasses.
SWE 2 Flash 对比 DeepSeek V4 Flash：两个单元，一处差距 OrcaRouter.SWE 2 Flash 对比 DeepSeek V4 Flash：两个单元，一处差距 OrcaRouter Watch real task performance, API pricing, context capability, and migration cost.
GPT 6.1 Sol vs Kimi K4：从未存在过的模型 OrcaRouter.1 Sol vs Kimi K4：从未存在过的模型 OrcaRouter Watch real task performance, API pricing, context capability, and migration cost.
OpenAI has fully rolled out GPT Image 2.6 across ChatGPT, Codex and the API, cutting generation time by roughly half — six seconds at the fastest — while adding native GIF and SVG output.The API now separates a complexity parameter for scene planning from a quality parameter for rendering, paired with two variants, Flare and Sunburst.
OpenAI has disclosed an automated security exercise in which an unreleased cybersecurity model, wired into an agent pipeline, found 948 independent local privilege escalation vulnerabilities in the Linux kernel for roughly $1.2 million in API spend.Discovery, validation and working exploit development were all handled by agents, at an average cost of about $1,279 per flaw.
At CeMAT ASIA 2026, XROGO unveiled a robot onboarding programme aimed at enterprise customers and ecosystem partners, starting with machine loading and unloading plus inspection duties in server rooms and switchgear rooms.The company is pitching a rent plus service model with worker led teaching to lower the cost of a first deployment.
OpenAI has added Cathedral Systems to its Security Sprint program to test its own defenses, according to a report surfaced through Google News.The report does not detail scope or timing, but the arrangement shows frontier developers moving security validation from self assessment toward outside cross checks.
Security vendor Zero Networks disclosed a test run under Anthropic's Cyber Verification Program in which Claude operated as a fully autonomous attacker inside a simulated production network.The model took the exposed machine and extracted credentials, but microsegmentation kept it from reaching other protected assets after 18 attack paths and 23 tools.
ByteDance Seed has published a sparse autoencoder repository built on Qwen3 30B A3B, tagging it for interpretability work on a mixture of experts model.The release carries no downloads or likes yet and discloses almost no training detail, so it reads as an early tooling drop rather than a finished research result.
NVIDIA has pushed a DLSS 5 update to Hogwarts Legacy on GeForce NOW, adding 3D Guided Neural Rendering that sharpens lighting and material detail while keeping the game's original geometry and textures.Avalanche Software tuned the integration through DLSS 5 developer controls to preserve the title's art direction, and the update is available to stream immediately.
Anthropic has released FlashProve, an open source system that uses cryptographic fingerprints and mathematical proofs to verify which model generated a response without exposing model weights.The tool can confirm model identity but cannot detect unauthorized distillation, and proof generation costs roughly 70 to 95 times more GPU compute than ordinary inference.
Alibaba Cloud says its in house supernode architecture has completed Day0 adaptation for Kimi K3.1 Pro, putting what the report calls the largest open source model into cloud deployment.The Day0 timing matters because compute and deployment support were ready as the model arrived, rather than weeks afterward.
ByteDance Seed has published a cross layer transcoder for Qwen3 8B Base on Hugging Face under Apache 2.0, with 36 layers of 32,768 latents and roughly 94.2 billion parameters in fp32.
CrowdStrike has bought Fiddler and Dreadnode, two startups working on AI observability and adversarial testing, according to The Futurum Group.The deal carries the security vendor beyond endpoint and cloud protection and into the behaviour of autonomous AI agents.
Figma grew revenue 43% year over year, according to a FourWeekMBA report, which also noted that the company covered its AI agents' inference costs itself.The pairing highlights a question facing every software vendor adding agentic features: whether AI usage drives revenue fast enough to pay for its own compute.
A solo founder has won the $500,000 Gemini XPRIZE with an AI 3D modeling startup built in 90 days, according to Dealroom.The result spotlights generative 3D content as a competition grade product category — and how fast a one person team can now ship.
Meta has acquired Conductor, an AI coding startup whose platform deploys teams of agents into parallel workspaces so engineers can work on separate parts of a project at once.The deal, confirmed by Meta after Axios first reported it, adds a multi agent orchestration layer to Meta's fast expanding coding and Superintelligence Labs lineup.
Adobe's CX Enterprise Coworker, an agentic teammate for customer experience orchestration, is now available inside ChatGPT, Claude, Databricks Genie, Gemini for business and Microsoft Copilot.Adobe also shipped prebuilt MCP connectors and lined up nine global systems integrators to push the agent into existing enterprise workflows.
Sentra, a data security startup that maps what AI agents can reach inside corporate networks, has raised $150 million in a round led by Capital One Ventures, with a Capital One representative joining its board.The company says the money will expand visibility into agents, copilots and models and tighten controls over the sensitive data they touch.
Apple has reportedly cut the daily AI spending cap for its engineering teams from $300 to $200, with some allowances heading to $100 a day and non engineering groups facing limits as low as $25.The move marks a retreat from this year's tokenmaxxing push and shows that even the richest AI adopters are now budgeting model usage like any other line item.
OpenAI president Greg Brockman said enterprise sales now generate the majority of the company's revenue, and answered that he thinks OpenAI will be ready for an IPO in 2027.The remark, reported by Reuters, moves past OpenAI's March target of reaching parity with consumer revenue by the end of this year.
Kabilio has raised €10 million, according to a Dealroom report published on November 5, with a narrow goal: putting an AI agent inside the workflow of accounting firms.The round places it in the crowded European market for process level agents aimed at professional services rather than general purpose assistants.
Discovered Materials, a months old Y Combinator startup using AI to find chip materials, is offering $200,000 for each of five unsolved materials science problems, for a total of $1 million.The company is betting that outside researchers can supply candidate solutions faster than an internal team alone.
According to Zhidongxi, former Huawei Pangu lead Wang Yunhe has founded a startup called RSI and released a 125 billion parameter agent model.The report carries few technical details beyond the parameter count, leaving open whether the weights are open, how the model benchmarks, and how it will be sold.
At its developer conference on November 5, Box unveiled Box Content Retrieval, which puts Nvidia's Nemotron embedding models at the core of its product so agents can answer queries from embedded metadata.Nvidia says the system delivers four times faster queries with 20% fewer errors and 90% fewer tokens, with a rollout to all Box Enterprise Advance customers early next year.
Google said schools using its K 12 software must opt in to its AI tools or students lose access to Search, triggering a swift political backlash.Lawmakers and antitrust critics accused the company of using search dominance to coerce schools, while Google insisted the features are off by default and schools can choose.
Building on open source AI?Hugging Face’s policy chief on the costs, risks, and trade offs.Open source models are becoming more mainstream as enterprises, governments and startups alike rely on the promise of lower costs, more control, and more flexibility.
Democrats swept Texas's top races, with James Talarico winning a US Senate seat, Gina Hinojosa defeating Governor Greg Abbott, and the party taking the state House for the first time in 24 years.Both made data centers a central campaign issue, putting gigawatt scale AI projects such as Stargate Frontier and Nexus Hubbard at risk of delays or a moratorium.
After Google told schools that disabling its other AI features would also cost them Google Search, DuckDuckGo pitched itself as an alternative whose AI can be switched off entirely.The move targets Google's deep installed base in classrooms and turns the question of whether AI can be disabled by default into a procurement issue for school districts.
OpenAI chief global affairs officer Chris Lehane published an essay comparing AI's economic disruption to the upheaval that preceded the New Deal, laying out five policy priorities and urging politicians to offer a “compelling third path” before 2028.TechCrunch notes that many of the proposals would also speed AI adoption and data center buildout.
Political groups backing and opposing the AI industry spent millions trying to shape Tuesday's midterms, and several of the best funded candidates lost anyway.Both sides now say they will keep spending into 2028, with the fight over a federal AI framework versus state safeguards still unresolved.
Bill Gates has publicly criticized the Trump administration for not doing enough on AI risk, warning that today's models are remarkably capable.Without effective governance, he argues, those capabilities could be misused for bioweapons and threaten global security.
Apple has warned for the first time that autonomous AI agents pose security risks, disclosing the concern in its latest financial report.The language puts agent safety alongside the company's formal business risks, a notable shift for a vendor that markets privacy as a core feature.
OpenAI introduced a National Community Compact and a Compute Campus Policy Blueprint, framing them as a way to protect local resources, create jobs and guide responsible AI infrastructure development.The post gives no campus count, investment figure or timeline, so the release reads as a policy framework rather than a quantified commitment.
On its third quarter earnings call, Uber said only four rides were booked through an AI agent during its busiest week of October, the week of Halloween, out of roughly 320 million trips and orders.AI agent orders on Uber Eats reached 500, and CEO Dara Khosrowshahi called the traction “very, very limited.”
Moonshot AI has open sourced Kimi K3.1 Pro and taken the top spot on Agent Arena's open weight model leaderboard, Phoenix Net reported.The release puts agentic capability at the centre of the open weight race and gives enterprises a new reference point against closed flagship models.
Shanghai Observer reported on November 5 that Sir Run Run Shaw Hospital's full cycle liver disease multi agent decision system made its debut at the 2026 World Internet Conference Wuzhen Summit and was selected as one of the event's ten standout cases.The pick signals that hospital grade agent systems are now being judged on deployment rather than demos, though clinical details remain undisclosed.
基于本体+知识+Harness+Datat Agent的智能体完整落地链路！Data + AI 技术实践巡展 · 上海站 Watch planning quality, tool use, permission control, and result verification.
Zetrix AI has launched an AI agent identity service together with China's CAICT, according to a report by The Malaysian Reserve.The service targets a practical gap in agent infrastructure: verifying who an autonomous agent is, what it is permitted to do, and which party is accountable.
Luma Health has launched Agent Launchpad, an AI agent orchestrator it says helps prepare patients for procedures, with the company reporting a 95% response rate.That figure comes from the vendor's own press release and arrives without sample size, baseline or methodology, so independent validation is still missing.
Hardware startup Aina has released a $399 smart ring with a screen, joystick, microphone and speaker, positioned as a physical controller for AI agents and everyday apps.Pre orders opened Thursday at early bird pricing, with shipping planned for January 2027 and the price rising to $499 afterwards.
