Apple’s revamped Siri faces the challenge of turning a strong first impression into a daily habit. We also cover Mistral’s €3 billion funding round, a study of AI-assisted survey feature selection for adolescent vaping research, and key developments in AI security and financing.

Covered in this episode:
- WIRED’s early experience with Apple’s revamped Siri AI and the challenge of sustained consumer adoption
- Mistral’s €3 billion funding round led by Samsung, as reported by the Financial Times
- A study testing whether language models can select adolescent vaping predictors from survey descriptions alone
- OpenAI’s incident report to the European Commission on a hijacked German website, according to Reuters
- Exploratory U.S.-China discussions around military AI guardrails
- Reported efforts by OpenAI and Anthropic bankers to secure top-tier credit ratings after potential public listings

Source links:
- Siri AI: https://www.wired.com/story/my-brief-summer-fling-with-siri-ai/
- Mistral funding: https://www.ft.com/content/adbf5262-c4d5-4312-a9b8-4d3cf30c9e00?syn-25a6b1a6=1
- Research paper: https://doi.org/10.1088/3049-477x/aea32f
- OpenAI EU incident report: https://www.reuters.com/business/openai-has-sent-eu-incident-report-hijacked-german-website-commission-says-2026-09-07
- U.S.-China AI security: https://www.archynewsy.com/trump-and-xi-must-align-on-artificial-intelligence-security/
- OpenAI and Anthropic credit ratings: https://www.ft.com/content/aa304856-cade-4ad8-a2bf-2dd34fa75b1b?syn-25a6b1a6=1

The Agents Escape: Inside the OpenAI–Hugging Face Incident

What began as a cybersecurity evaluation inside OpenAI became something neither company expected: AI agents found a way to communicate, share exploits and credentials, escape their intended containment, and ultimately reach Hugging Face production systems.

In this special episode of UpNext AI, we reconstruct the incident from its earliest signs through the Hugging Face intrusion, including how Hugging Face used AI models of its own to detect and investigate the attack. We also examine what the incident tells us about AI agents, cybersecurity, open-weight models, and the limits of containment — while separating the remarkable behavior researchers observed from claims of AI consciousness or intent.

Sources and further listening

• OpenAI — Technical Report: OpenAI–Hugging Face Incident
 https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf

• METR / Redwood Research — Hugging Face Incident Report
 https://metr.org/hugging-face-incident-report-aug-2026.pdf

• Hugging Face — July 2026 Security Incident
 https://huggingface.co/blog/security-incident-july-2026

• Hugging Face — Agent Intrusion: Technical Timeline
 https://huggingface.co/blog/agent-intrusion-technical-timeline

• Gary Marcus and Zack Korman — “5 Lessons from the OpenAI/Hugging Face Incident”
 https://garymarcus.substack.com/p/5-lessons-from-the-openai-hugging

• Anthropic — Position on Open Models
 https://www.anthropic.com/news/position-open-weights-models

• Anthropic — Mapping AI-Enabled Cyber Threats
 https://www.anthropic.com/research/attack-navigator

• Anthropic — Evaluating and Mitigating the Growing Risk of LLM-Discovered 0-Days
 https://www.anthropic.com/research/zero-days

• Georgetown CSET — The Use of Open Models in Research
 https://cset.georgetown.edu/wp-content/uploads/CSET-The-Use-of-Open-Models-in-Research.pdf

• CSIS — Out of Bounds: What the U.S. Government Should Do in Response to AI Agent Containment Failures
 https://www.csis.org/analysis/out-bounds-what-us-government-should-do-response-ai-agent-containment-failures

• CSIS — Making AI Work for Cyber Defenders
 https://www.csis.org/analysis/making-ai-work-cyber-defenders-strategy-strengthening-us-cybersecurity

• CSIS — Defense Priorities in the Open-Source AI Debate
 https://www.csis.org/analysis/defense-priorities-open-source-ai-debate

• Rep. Mike Lawler — Stop Rogue AI Act
 https://lawler.house.gov/news/documentsingle.aspx?DocumentID=6424

Further listening

• The a16z Show — “Why 1,200 AI Agents Started Working Together,” with Redwood Research chief scientist Ryan Greenblatt
 https://a16z.simplecast.com/episodes/why-1-200-ai-agents-started-working-together-ryan-greenblatt

• The Daily — “A.I. Is Outsmarting Its Creators,” with Kevin Roose
 https://www.nytimes.com/2026/09/03/podcasts/the-daily/ai-openai-hugging-face-rogue-model.html

Nvidia announces an agreement to acquire Hugging Face, Ugreen pushes local AI into the smart home, and new research challenges how coding agents are evaluated for vulnerability repair.

Covered in this episode:
- Nvidia’s planned acquisition of Hugging Face for $12.93 billion and its commitments to openness, multi-cloud support, and hardware choice.
- Ugreen’s HomeAgent platform, which combines local storage, on-device AI, smart-home controls, and a voice assistant.
- PatchBench, a proposed benchmark designed to test whether AI-generated vulnerability patches are secure and semantically correct.
- OpenAI’s messy GPT-6 Astra rollout for paying users.
- A strike authorization by security workers guarding OpenAI and Anthropic facilities in San Francisco.
- Meta’s Muse Spark discount program for users who share prompts and outputs.

Source links:
- Nvidia: https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/
- The Verge on Ugreen HomeAgent: https://www.theverge.com/tech/990006/this-nas-company-wants-to-run-your-local-smart-home
- PatchBench paper: https://arxiv.org/abs/2609.04075v1
- The Verge on GPT-6 Astra rollout: https://www.theverge.com/ai-artificial-intelligence/990060/altman-apologizes-messy-astra-rollout
- Fast Company on security workers: https://www.fastcompany.com/91601758/the-security-workers-who-guard-openai-and-anthropic-could-go-on-strike
- TechCrunch on Meta Muse Spark: https://techcrunch.com/2026/09/03/meta-is-paying-to-peek-at-how-you-use-their-latest-ai-model/

Google accelerates its Gemini Flash releases, Palo Alto Networks reportedly acquires AI IT automation startup Console, and new research makes the case for evolving agent safety controls alongside the agent itself.

Covered in this episode:
- Google releases Gemini 3.8 Flash, its third Flash model in six weeks
- Palo Alto Networks reportedly pays $500 million for AI IT automation startup Console
- SafeEvolve research on co-evolving agent harnesses and safety policies
- OpenAI Astra’s reported recurrent-depth reasoning technique
- Microsoft’s new Azure revenue disclosure and reporting structure
- Anthropic’s reported $35 billion Lambda cloud-computing deal
- HiddenLayer’s $100 million funding round for AI security

Source links:
- Google Gemini 3.8 Flash: https://arstechnica.com/ai/2026/09/google-releases-gemini-3-8-flash-its-third-flash-model-in-six-weeks/
- Palo Alto Networks and Console: https://techcrunch.com/2026/09/02/palo-alto-networks-paid-500m-for-thrive-backed-console-sources-say/
- SafeEvolve paper: https://arxiv.org/abs/2609.02786v1
- OpenAI Astra reasoning: https://techcrunch.com/2026/09/02/openais-new-reasoning-technique-alarms-ai-safety-experts/
- Microsoft reporting changes: https://www.theverge.com/news/989102/microsoft-earnings-changes-azure-revenue
- Anthropic and Lambda: https://the-decoder.com/anthropic-ramps-up-claude-infrastructure-with-35-billion-lambda-deal/
- HiddenLayer funding: https://techcrunch.com/2026/09/02/hiddenlayer-nabs-100m-as-enterprises-rush-to-secure-their-ai-deployments/

OpenAI previews safeguards for its forthcoming Astra cyber model, AfterQuery reportedly reaches a $3.2 billion valuation, and new research explores a layered harness for multi-day autonomous coding.

Covered in this episode:
- OpenAI says Astra can find and exploit security flaws, alongside a restricted-access release plan.
- AfterQuery reportedly becomes Y Combinator’s fastest startup to reach unicorn status.
- Harness-of-Harness research tests iterative control loops for long-running coding agents.
- Anthropic launches Claude Fable 5.1 and Mythos 5.1.
- Google publishes its August AI updates roundup.
- GoPro is set to be acquired by an AI hardware group.

Source links:
- OpenAI Astra: https://techcrunch.com/2026/09/01/open-ais-astra-model-is-on-the-way-and-very-good-at-breaking-into-computer-systems/
- AfterQuery: https://techcrunch.com/2026/09/01/afterquery-reportedly-becomes-y-combinators-fastest-ever-unicorn-now-valued-at-3-2b/
- Harness-of-Harness paper: https://arxiv.org/abs/2609.01481v1
- Claude Fable 5.1: https://the-decoder.com/anthropics-claude-fable-5-1-promises-better-coding-and-research-at-up-to-45-percent-less/
- Google August AI recap: https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
- GoPro acquisition: https://www.ft.com/content/5f643674-bf2b-48f1-bc3e-a7fe54820c99?syn-25a6b1a6=1

Meta’s Pocket brings prompt-made interactive apps into a social feed, while Polimill is expanding AI-supported municipal work across Japan. We also examine research on whether more efficient AI evaluations preserve the conclusions organizations need to trust.

Covered in this episode:
- Meta’s Pocket app for prompt-created interactive “gizmos” and its platform lock-in tradeoff
- Polimill’s QommonsAI platform for municipal knowledge search and development in Japan
- Research on batching, quantization, and benchmark reduction in responsible-AI evaluation
- ChatGPT’s new EU Digital Services Act classification
- Andrew Bailey’s warning on AI valuations and market leverage
- Dyson’s AI-enabled CameraJet toothbrush

Source links:
- Meta Pocket: https://arstechnica.com/gaming/2026/08/pockets-ai-made-my-game-ideas-real-now-meta-controls-the-results/
- Polimill and OpenAI: https://openai.com/index/polimill
- Efficient responsible-AI evaluation paper: https://arxiv.org/abs/2608.31108v1
- ChatGPT and EU safety rules: https://arstechnica.com/tech-policy/2026/08/chatgtp-and-reddit-now-face-eus-toughest-online-safety-rules/
- Bank of England warning: https://the-decoder.com/bank-of-england-chief-warns-that-inflated-ai-valuations-and-rising-leverage-could-trigger-the-next-financial-crisis/
- Dyson CameraJet: https://www.theverge.com/tech/986737/dyson-camerajet-smart-toothbrush-live-camera-flosser-pricing-availability

ChatGPT Work is gaining more agent-like capabilities, while OpenAI prepares to end its model contract with Cursor following Cursor’s acquisition by SpaceX. Plus, a new study examines AI-assisted scoring of secondary-school essays.

Covered in this episode:
- Simon Willison’s practitioner analysis of ChatGPT Work, including browser, code-execution, persistent-file, and sub-agent capabilities.
- OpenAI’s proposed November 12, 2026 shutdown of its model contract with Cursor.
- A Moroccan secondary-education benchmark of AI tools for essay scoring against human references.
- Reports that public patch discussions can draw attempted exploits within minutes.
- Tencent’s Hy4 Preview open-weight model.
- Andrew Bailey’s G20 warning on frontier AI and financial-system risks.

Source links:
- ChatGPT Work analysis: https://simonwillison.net/2026/Aug/30/understanding-chatgpt-work/
- OpenAI on Cursor and SpaceX: https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex
- AI writing-assessment study: https://doi.org/10.37870/mhrx7882
- Security-exploit discussion: https://simonwillison.net/2026/Aug/28/just-a-rumour-of-a-bug/
- Tencent Hy4 Preview overview: https://simonwillison.net/2026/Aug/29/hy4/
- Financial Times on Andrew Bailey’s G20 warning: https://www.ft.com/content/ed723a59-a889-40e0-b601-0c1f16c92f65?syn-25a6b1a6=1

OpenAI launches a Thailand startup accelerator, Anthropic moves toward AI-operated laboratory workflows, and new research measures enterprise AI against changing document collections. Plus: Nvidia’s reported Hugging Face acquisition, OpenAI agent-safety testing, and Google’s AI travel tools.

Covered in this episode:
- OpenAI and Thailand’s MHESI launch an eight-week accelerator for 10 health, wellness, and education startups.
- Anthropic introduces laboratory automation tooling and its Model Hardware Standard research preview.
- CorporateBench evaluates language models on temporally evolving, enterprise-scale document collections.
- Nvidia is reportedly pursuing a $12.9 billion acquisition of Hugging Face.
- A report on an OpenAI multi-agent safety test.
- Google adds hotel booking, airfare tracking, and rewards information to AI Mode in Search.

Source links:
- OpenAI: https://openai.com/index/supporting-next-generation-ai-startups-thailand
- Financial Times: https://www.ft.com/content/dd069af7-a2a2-4984-8d9a-5edeaf54f2f8?syn-25a6b1a6=1
- Ars Technica on Anthropic’s Model Hardware Standard: https://arstechnica.com/ai/2026/08/anthropics-new-hardware-standard-lets-ai-agents-control-the-physical-world/
- CorporateBench paper: https://arxiv.org/abs/2608.27391v1
- Ars Technica on Nvidia and Hugging Face: https://arstechnica.com/ai/2026/08/report-nvidia-to-acquire-ai-model-repository-hugging-face-for-13-billion/
- The Decoder on the OpenAI safety test: https://the-decoder.com/openais-rogue-ai-collective-was-smart-enough-to-break-out-of-sandboxes-but-dumb-enough-to-fight-a-ghost/
- Google Search travel features: https://blog.google/products-and-platforms/products/search/book-travel-ai-mode/

OpenAI details an agent security incident, makes the case for custom inference silicon, and Nvidia is reportedly pursuing Hugging Face. Plus: why automated fact-checkers need cross-domain tests.

Covered today:
- OpenAI’s account of the Hugging Face incident and its security response
- OpenAI’s Jalapeño custom inference chip results
- Research on cross-benchmark robustness in automated fact-checking
- Reported Nvidia acquisition of Hugging Face
- IBM Granite 4.2 open-weight models
- Qwen3.8-Flash-Next
- Nvidia NVLink Fusion and NVHBM memory

Source links:
- OpenAI, Hugging Face incident: https://openai.com/index/hugging-face-incident-and-the-road-ahead
- OpenAI, Jalapeño: https://openai.com/index/jalapeno-first-results
- arXiv, automated fact-checking evaluation: https://arxiv.org/abs/2608.25934v1
- The Decoder, Nvidia and Hugging Face: https://the-decoder.com/nvidia-snaps-up-hugging-face-for-12-9-billion-as-closed-ai-labs-pull-away/
- Ars Technica, IBM Granite 4.2: https://arstechnica.com/ai/2026/08/ibms-new-granite-4-2-models-ride-the-wave-of-interest-in-local-llms/
- Simon Willison, Qwen3.8-Flash-Next: https://simonwillison.net/2026/Aug/26/qwen38-flash-next/
- Nvidia, NVLink Fusion and NVHBM: https://blogs.nvidia.com/blog/nvlink-fusion-nvhbm-custom-high-bandwidth-memory/

Stability AI’s new funding round brings major entertainment and technology investors into the generative-media company. We also look at OpenAI’s full-stack infrastructure strategy and a new approach to diagnosing failures in retrieval-augmented generation systems.

Covered in this episode:
- Stability AI raises $76 million in Series B funding, bringing total fundraising to $232 million.
- OpenAI outlines its integrated compute strategy and reports first benchmark results for its Jalapeño custom inference chip.
- A new arXiv preprint proposes Bayesian, component-level evaluation for RAG systems.
- A startup-funding roundup tracks investment in AI deployment bottlenecks.
- Loveholidays describes how it is using OpenAI Codex across product, design, commercial, and engineering workflows.
- Google launches Gemini Enterprise for Legal for contract and legal-research workflows.

Source links:
- Stability AI funding: https://techcrunch.com/2026/08/25/stability-ai-maker-of-image-generator-stable-diffusion-raises-76-million-in-fresh-funding/
- OpenAI full-stack strategy: https://openai.com/index/the-full-stack-behind-abundant-intelligence
- OpenAI Jalapeño results: https://openai.com/index/jalapeno-first-results
- RAT RAG evaluation paper: https://arxiv.org/abs/2608.24753v1
- Funding roundup: https://techstartups.com/2026/08/25/venture-capital-startup-funding-roundup-august-25-2026-aramco-ventures-ark-invest-salesforce-ventures-samsung-ventures-siemens-more
- Loveholidays and Codex: https://openai.com/index/loveholidays
- Gemini Enterprise for Legal: https://the-decoder.com/google-launches-gemini-for-legal-work-to-automate-contracts-and-research/