UpNext AI

Amplitude’s Wave connects OpenAI-powered product analysis with coding agents and impact measurement; Hugging Face CEO Clément Delangue responds to Jensen Huang’s departure amid Nvidia’s pending acquisition. Also covered: Nectar’s Creator Agent, Metal Sandbox for Apple Silicon, Anthropic’s Claude compute strategy, and JioStar chief architect Vijay Seshadri on AI discovery.Amplitude’s Wave proposes product changes, hands plans to coding agents, and measures results. [Source](https://openai.com/business/customer-stories/amplitude)Delangue pledges continued support for open-source AI as Nvidia’s planned Hugging Face acquisition remains pending. [Source](https://techcrunch.com/2026/11/12/hugging-face-ceo-vows-to-keep-fighting-for-open-source-ai-after-jensen-huangs-exit/)Nectar launches Creator Agent for discovering creators, managing outreach, and tracking partnerships. [Source](https://techcrunch.com/2026/11/12/nectar-launches-an-ai-agent-to-automate-creator-discovery-and-outreach-for-brands/)Metal Sandbox researchers test scheduling GPU coding agents across Apple Silicon machines. [Paper](https://arxiv.org/abs/2611.17294v1)Anthropic describes its multi-chip approach to training and serving Claude in an IPO roadshow video. [Source](https://simonwillison.net/2026/Nov/11/anthropic-ipo-roadshow/)Vijay Seshadri discusses JioStar’s AI discovery work and production considerations. [Interview](https://openai.com/index/jiostar-vijay-seshadri)

The Information reports heightened concern about AI’s pace, while TSMC plans investments in OpenAI and Anthropic. Also covered: OpenAI’s children’s safety research commitment, the Pyrite protein–ligand pose-ranking benchmark, Alibaba’s chip-access-related share decline, Google DeepMind’s AlphaGenome 2, and inference startups buying chips.The Information on AI acceleration and safety researchers’ concerns: https://www.theinformation.com/articles/anthropic-others-warn-ai-accelerationTSMC’s planned OpenAI and Anthropic investments: https://arstechnica.com/ai/2026/11/taiwans-chipmaking-giant-tsmc-to-invest-1b-in-anthropic-openai/OpenAI’s commitment to children’s AI safety research: https://openai.com/index/children-first-declarationPyrite’s benchmark for ranking candidate protein–ligand binding poses: https://arxiv.org/abs/2611.14511v1Alibaba’s decline amid advanced-chip access concerns: https://www.bloomberg.com/news/articles/2026-11-12/alibaba-shares-drop-7-as-nvidia-saga-triggers-chip-access-fearsGoogle DeepMind’s AlphaGenome 2: https://www.theverge.com/ai-artificial-intelligence/1031524/google-deepmind-alphagenome-2-genetic-ai-modelInference startups beginning to buy their own chips: https://www.theinformation.com/articles/ai-inference-startups-quietly-buying-chips

Microsoft researchers describe Lazarus, a hidden-trigger backdoor for AI agents, while Google DeepMind introduces Gemini Robotics 2.5 and the ASIMOV-jailbreak safety benchmark. Also covered: Mosaic Intelligence Labs’ surgical catheter, a study of enterprise analytics agent evaluations, Nvidia chip routes through Singapore, Google’s Adobe integration, and CoreWeave’s planned chip-backed financing.Microsoft’s Lazarus research examines malicious edits to model weights and the importance of provenance and agent permissions.Google DeepMind’s Gemini Robotics 2.5 pairs dexterous-task demonstrations with an open safety benchmark.Mosaic Intelligence Labs has raised $47 million while developing a semi-autonomous surgical catheter.An enterprise-agent evaluation preprint compares trace-based judging, text-only judging, and fixed rules.The Financial Times examines a Singaporean chip deal in the context of Nvidia chips and China.Google outlines Adobe integrations for Drive and Gemini.CoreWeave plans a chip-backed debt deal and is using customer pre-payments for financing.Full source links:- Microsoft Research: https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents/- Ars Technica: https://arstechnica.com/google/2026/11/deepmind-reveals-gemini-robotics-2-5-and-open-safety-benchmark-asimov-jailbreak/- The Information: https://www.theinformation.com/articles/mosaic-raises-funds-robotic-surgery- Enterprise-agent evaluation preprint: https://arxiv.org/abs/2611.14493v1- Financial Times on Singapore and Nvidia chips: https://www.ft.com/content/52b5f9a8-346f-4dae-8a16-14d5d9eb1244?syn-25a6b1a6=1- Google on Adobe integration: https://blog.google/company-news/inside-google/company-announcements/adobe-google-partnership-integration/- Financial Times on CoreWeave: https://www.ft.com/content/fa5c5bdc-c2b0-4fab-8951-0526a4608773?syn-25a6b1a6=1

CodeRabbit CEO Harjot Gill discusses Codex-assisted software delivery and CodeRabbit Secure, while OpenAI details its Control Room for stopping research workloads and Meta introduces Muse Ember for repetitive agent tasks. Also covered: DeidSharp’s clinical-text de-identification evaluation, proposed Nvidia chip export controls, security executive Lisa Thiergart’s assessment of espionage risks, and a Claude Fable 5.2 prompting test.CodeRabbit describes shipping roughly twenty times more software with the same engineering team and explains how it reviews and secures agent-generated code. Source: https://openai.com/index/coderabbit-codex-daybreakOpenAI explains how Control Room tracks research jobs, denies inference requests for stopped jobs, and fails closed if its service goes down. Source: https://openai.com/index/control-roomMeta launches Muse Ember; Simon Willison examines its agent-task positioning, short-prompt pricing, and longer-prompt threshold. Source: https://simonwillison.net/2026/Nov/10/muse-ember/A DeidSharp preprint tests layered, locally hosted detection of protected health information in clinical text. Source: https://arxiv.org/abs/2611.12709v1The Information reports renewed consideration of restrictions involving advanced Nvidia chips and remote access from China. Source: https://www.theinformation.com/articles/nvidias-next-headache-new-push-export-controls-wake-spy-scandalSecurity Level 5 CEO Lisa Thiergart discusses suspected espionage risks at U.S. AI companies. Source: https://www.theinformation.com/articles/likely-spies-ai-companies-says-security-executive-lisa-thiergartAnthropic’s small coding-task tests show how an ownership instruction changes Claude Fable 5.2’s tendency to merge repeated code. Source: https://simonwillison.net/2026/Nov/10/prompting-claude-fable-52/

NVIDIA and Korea’s Ministry of Science and ICT are expanding work on sovereign AI infrastructure, NVLink Fusion and Korean models, while NVIDIA names investments involving Point2 Technology and plans for Motif Technologies, RLWRLD and Upstage. Also covered: Anthropic’s reported gas-fired power deal, Yoko Hasebe’s proposed autonomous-system controls, OpenAI’s Hacktron partnership, Simon Willison’s AI pull-request workflow, and the reported Crusoe–Google data center sale.NVIDIA and Korea expand collaboration on chips, infrastructure, inference software and skills. https://blogs.nvidia.com/blog/korea-sovereign-ai-collaboration/NVIDIA announces a Point2 investment and plans further Korean startup investments. https://blogs.nvidia.com/blog/korean-startup-investments-msit/The Financial Times reports Anthropic’s gas-fired power deal for data centers. https://www.ft.com/content/41bf5c6a-9735-4f1a-a8e3-2bcccad58ba0?syn-25a6b1a6=1Yoko Hasebe proposes a governance architecture separating AI reasoning from physical-system controls. https://doi.org/10.5281/zenodo.24041600OpenAI partners with Hacktron to test vulnerabilities and third-party connections. https://openai.com/index/partnering-with-hacktronSimon Willison shares a way to make agent-written pull requests easier to review. https://simonwillison.net/2026/Nov/9/details-summary/TechCrunch reports Crusoe’s sale of a Wyoming data center project to Google. https://techcrunch.com/2026/11/06/crusoe-sells-troubled-wyoming-data-center-project-to-google/

OpenAI’s GPT-6 Cyber pairs vulnerability research with authorized-scope safeguards, while an npj Sustainable Agriculture review examines agricultural agents and the SecureLLM study tests compact models against cybersecurity archives. XStack measures PyTorch-to-JAX repository migration; Brian Chesky’s AI startup, Dropbox’s Tigris Data acquisition, and Claude’s Plaid-connected personal finance feature round out the episode.OpenAI reports GPT-6 Cyber benchmark gains, Firefox vulnerability findings, Linux kernel exploit work, and Daybreak Red safeguards.A review maps potential uses and deployment challenges for AI agents in crop management, agricultural research, and robotics.SecureLLM researchers test continued training and retrieval using archived cybersecurity records, finding that lower training loss did not translate into stronger answers on their measures.XStack tests agents migrating complete machine-learning repositories from PyTorch to JAX.The Information reports on Airbnb founder Brian Chesky’s self-funded AI startup and Airbnb’s potential investment and licensing arrangement.Dropbox’s acquisition of Tigris Data is expected to support DBX Agent Cloud.Anthropic is launching read-only bank-account connections for Claude through Plaid.Sources:- OpenAI: https://openai.com/index/introducing-gpt-6-cyber- TechCrunch on the cybersecurity launch: https://techcrunch.com/2026/11/05/openai-released-a-new-cybersecurity-model-is-that-enough/- npj Sustainable Agriculture: https://www.nature.com/articles/s44264-026-00199-w- SecureLLM study: https://doi.org/10.21203/rs.3.rs-11209286/v1- XStack paper: https://arxiv.org/abs/2611.08472v1- The Information on Chesky’s startup: https://www.theinformation.com/articles/airbnb-reveals-new-details-cheskys-secretive-ai-startup- TechCrunch on Dropbox and Tigris Data: https://techcrunch.com/2026/11/05/dropbox-acquires-tigris-data-for-350m-to-bolster-its-ai-iinfrastructure/- The Information on Claude personal finance: https://www.theinformation.com/briefings/anthropic-launches-personal-finance-feature-claude

Realtime-Bench tests time-sensitive AI agents, while Google DeepMind previews Genie 4; OpenAI is acquiring Pierre Computer Company for Codex and offering GPT-5.6 Cyber through Daybreak Red on Amazon Bedrock in AWS GovCloud. Also covered: an arXiv study of text-embedding compression, Cyera’s discovery of unnoticed agents, Palo Alto Networks’ reported talks with Clutch Security, and DuckDuckGo’s pitch to schools.Realtime-Bench compares agent performance with running and paused clocks; Google DeepMind previews changes to interactive Genie 4 simulations. Source: https://importai.substack.com/p/import-ai-479-spacewar-low-latencyOpenAI plans to acquire Pierre Computer Company, whose Trees and Diffs libraries are used in Codex. Source: https://openai.com/index/openai-to-acquire-pierreDaybreak Red brings GPT-5.6 Cyber to approved customers through Amazon Bedrock in AWS GovCloud (US-East). Source: https://openai.com/index/daybreak-red-aws-govcloudResearchers benchmark shorter text embeddings and lower-precision storage, reporting a substantial storage saving with a task-dependent effectiveness trade-off. Source: https://arxiv.org/abs/2611.06859v1Cyera found previously unnoticed AI agents in a financial company’s sensitive data, according to The Information. Source: https://www.theinformation.com/articles/security-companies-trying-get-ai-agents-controlPalo Alto Networks is reportedly in talks to buy Clutch Security; no deal has been finalized. Source: https://www.theinformation.com/articles/palo-alto-networks-talks-pay-500-million-ai-security-startupDuckDuckGo is pitching schools on search and options to disable AI features, TechCrunch reports. Source: https://techcrunch.com/2026/11/04/duckduckgo-pitches-itself-to-schools-amid-backlash-over-googles-ai-policy/

Ted Schlein’s Ballistic Ventures is seeking new cybersecurity funds as CrowdStrike plans acquisitions of Fiddler AI and Dreadnode; Cisco describes its security work with OpenAI, while a Nature Aging paper introduces Quest2Dx for Parkinson’s questionnaire screening. Also covered: the DEP language-model evaluation toolkit, Anthropic’s FlashProve and Message Threads, Google’s school AI controls, and The Information’s reporting on OpenAI, Nvidia and chip sales to China.Ballistic Ventures targets $490 million across two funds; CrowdStrike plans to buy two AI security startups.Cisco and OpenAI describe network-informed security work and plans for the Daybreak Defense Network.A Nature Aging paper describes Quest2Dx, a language-model approach to questionnaire-based Parkinson’s screening.DEP proposes a shared protocol and toolkit for language-model evaluations.Anthropic is developing FlashProve to verify model identity.Google’s school AI controls come with an AI-enhanced search requirement, according to The Information.Anthropic’s Message Threads aims to reduce repeated conversation-history transfers.The Information reports on Nvidia’s influence over OpenAI’s public position on chip sales to China.Sources:- Ballistic Ventures: https://techcrunch.com/2026/11/03/ted-schleins-ballistic-ventures-seeks-490m-for-new-funds-amid-ai-driven-cybersecurity-boom/- CrowdStrike acquisitions: https://techcrunch.com/2026/11/03/crowdstrike-to-shell-out-nearly-500m-for-2-ai-security-startups-as-the-market-for-securing-agents-takes-off/- Cisco and OpenAI: https://openai.com/index/cisco-cyber- Quest2Dx study: https://www.nature.com/articles/s43587-026-01260-6- DEP paper: https://doi.org/10.1145/3799682.3840262- FlashProve: https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/- Google schools: https://www.theinformation.com/briefings/google-gives-schools-ai-control-requires-ai-search- Message Threads: https://the-decoder.com/anthropics-message-threads-cut-time-to-first-token-by-35-percent-for-long-conversations/- Nvidia and OpenAI: https://www.theinformation.com/articles/openai-afraid-jensen-huang

Legora CTO Jacob Lauritzen discusses confidential legal AI evaluations and Tabular Review; Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab report early weight-space programming experiments; and WebHarbor offers repeatable environments for browser-agent tests. Also covered: Anthropic’s Claude Code game-recreation rule, Simon Willison’s LLM release, OpenAI’s bug bounty, Marvell’s Lumentum deal, and Oracle’s AI infrastructure financing.Legora’s approach to evaluating legal agents without access to customer documents, and Lauritzen’s account of changing legal billing.Early experiments in reading and editing neural-network weights directly, including a reported repair to Qwen’s reasoning loops.Claude Code’s instruction against recreating specific existing games.WebHarbor’s website environments for training and evaluating graphical-interface agents.Simon Willison’s LLM tool release; OpenAI’s higher bug-bounty payouts; Marvell’s planned Lumentum acquisition; and Oracle’s debt and stock offerings.Sources:- Legora interview: https://latent.space/p/building-ai-for-lawyers-jacob-lauritzen-cto-of-legora/- Weight-space programming: https://rikiyat.github.io/blog/weight-space-programming/- Claude Code game rule: https://arstechnica.com/ai/2026/11/claudes-new-rules-against-game-cloning-wont-stop-the-ai-copying-boom/- WebHarbor: https://arxiv.org/abs/2611.03634v1- LLM release: https://simonwillison.net/2026/Nov/2/llm/- OpenAI bug bounty: https://the-decoder.com/bug-hunters-can-now-earn-up-to-172000-from-openai-as-company-ramps-up-security/- Marvell and Lumentum: https://www.theinformation.com/briefings/marvell-buy-lumentum-115-billion-all-stock-deal- Oracle offerings: https://www.theinformation.com/briefings/oracle-launches-20-billion-stock-bond-sales

Salesforce’s proposed Rune Technologies acquisition would extend Missionforce into field logistics for the U.S. military. Also covered: Nathan Lambert on open-weight models and cybersecurity, the Anthropic Claude extension incident in Cursor, a review of AI’s water footprint, and updates involving Claude watermarks, ChatGPT, Datasette Agent and empathetic chatbot prompts.Salesforce agrees to acquire AI logistics startup Rune Technologies in a reported deal worth $375 million.Nathan Lambert examines open-weight models, vulnerability discovery and the wider cyber response.A malicious version of Anthropic’s Claude extension was distributed through Cursor’s marketplace.A narrative review examines direct and indirect freshwater use across AI’s lifecycle.Anthropic expands Claude text watermarks, with a gap on older models.Parameter Labs finds a ChatGPT share-link vulnerability, which OpenAI closed.Datasette Agent’s underlying plug-in simplifies default-model setup.Researchers examine how empathetic prompts affect chatbot responses.Sources:- Salesforce and Rune: https://techcrunch.com/2026/10/30/salesforces-ma-spree-continues-with-a-reported-375m-deal-for-defense-tech-startup-rune/- Open models and cyber risk: https://www.interconnects.ai/p/was-i-wrong-about-open-models-and- Claude extension incident: https://arstechnica.com/security/2026/10/anthropic-vs-code-extension-on-cursor-backdoored-for-nearly-12-hours/- AI water-footprint review: https://doi.org/10.2166/wp.2026.233- Claude watermarks: https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/- ChatGPT vulnerability: https://simonwillison.net/2026/Oct/31/how-we-took-over-chatgpt-accounts/- Datasette Agent plug-in: https://simonwillison.net/2026/Oct/31/datasette-llm/- Empathetic chatbot prompts: https://the-decoder.com/empathetic-prompts-can-turn-ai-chatbots-into-seemingly-traumatized-conversational-partners/