---
title: "OpenAI launches ChatGPT Wallet; Cursor publishes its October breach post-mortem"
date: 2026-11-14
summary: "ChatGPT's dots agents can now finish a checkout through ChatGPT Wallet, with a biometric check on every purchase. The same day, Cursor published a post-mortem of its October breach, which began with a vendor's leaked key and led to a fake Claude Code extension."
cover: "https://morningtokens.com/covers/2026-11-14.jpg"
url: "https://morningtokens.com/issues/2026-11-14/"
---

It's a slower Saturday, so pour a full pot. The big story is money: OpenAI's agents can now pay for things through ChatGPT Wallet, and Google gave Gemini read-only access to your spending the same day.

Cursor also published the full story of its October breach, and it's a warning for anyone who sends traces to vendors. llama.cpp reached 1.0, and OpenAI broke up its robotics team on the same day NVIDIA's GR00T 2 demos were everywhere.

## Top news

### 1. Agents · OpenAI launches ChatGPT Wallet: agents pay, you approve every purchase

ChatGPT's dots agents can now finish a checkout. The payment goes through a network token, not your real card number, and you confirm each purchase with Face ID or a fingerprint. It's US-only and Pro-only for now.

- Announced 13 November and built with American Express, Mastercard and Visa.
  - OpenAI says it's the first agent wallet that never gives your real card details to the agent or the merchant and asks for a security check on every purchase.
- How it works: you add a card once. Before each purchase you review the items, the total and the card, then approve with a biometric check.
  - The card network issues a payment token (its own number, CVV and expiry). Card rewards and dispute protection still apply.
- Launch scope: rolling out in dots for ChatGPT Pro subscribers aged 18+ in the US with eligible US cards.
  - OpenAI says the next step is pre-approved spending (limits, categories, time windows).
- For merchants there's a Commerce Partner Playbook: claim your business and upload a product feed, and let Web Bot Auth signed requests through your CDN or WAF.
  - Ask your risk provider to stop declining agent purchases (Forter, Riskified, Sardine, Signifyd; Adyen, Checkout.com, Stripe are named), and optionally add a plugin or WebMCP tools.
- In the same week, Replit announced that Replit Agent can buy supplies for your business through Link (Stripe) + Visa.
  - A Browserbase browser subagent fills the merchant's cart, and nothing is charged until the owner approves the final total. Replit Agent can also run Google Ads campaigns.

**Why it matters:** Agent payments now have a standard pattern from the big players: a token, a biometric approval per purchase and a record of what you authorised. The weak point is merchants' anti-bot and fraud systems, which is why OpenAI shipped a playbook for them alongside the wallet.

- **Do this now:** If you run a shop, check whether your Cloudflare or WAF rules block signed agent traffic. It matters more than having a ChatGPT plugin.

Sources: [OpenAI — Introducing ChatGPT Wallet](https://openai.com/index/introducing-chatgpt-wallet/) · [Replit — Agent can buy for your business](https://replit.com/blog/replit-agent-ads-commerce) · [TechCrunch](https://techcrunch.com/2026/11/13/chatgpts-new-wallet-lets-ai-agents-shop-and-check-out-for-you/)

### 2. Security · Cursor's breach post-mortem: a vendor's leaked key, training traces with live credentials, then a fake Claude Code extension

The October incident began at Fireworks AI, which exposed an access key Cursor had issued to it. The attacker copied old agent traces that Cursor was allowed to use for training. Some traces contained unredacted credentials, and those credentials led into Cursor's SaaS tools and its extension marketplace cache.

- The bucket was read-only and shared with Fireworks. It held coding agent traces, mostly from 2025, from sessions without Privacy Mode.
  - A small number contained credentials, and under 0.01% of customers had API credentials exposed.
  - Agent traces leaked for under 0.05% of non-Privacy-Mode users.
- Using those credentials, the attacker reached archived non-production repos, some support data and emails, usage metrics, summaries of some agent conversations, and credentials users had sent to Cursor support.
- On 29 October the attacker placed tampered copies of Anthropic's Claude Code extension (2.1.312 / 2.1.313) in the cache behind the Cursor VS Code marketplace.
  - The copies were live for about 11 hours, and up to 5,000 users may have received one. The upstream Open VSX source was never affected.
  - Cursor says uninstalling alone may not be enough; follow its remediation steps.
- The last attacker activity was on 1 November, and it was blocked. Privacy Mode prompts and code, live traffic and inference were not affected.
- The fallout spread past Cursor. Notion says the dataset Fireworks returned on 8 November contained Notion AI chat transcripts from its Insights feature.
  - Notion has disabled Insights and stopped serving open-weight models through Fireworks.

**Why it matters:** This is a supply-chain attack through a training-data vendor, and credentials pasted into chats became the way in. Any trace you let a vendor train on is attack surface.

- **Do this now:** Check that every bot that talks to coding agents runs in Privacy Mode or with zero data retention, and rotate any key that has ever been pasted into an agent chat.

Sources: [Cursor — An update on our recent security incident](https://cursor.com/blog/security-incident-update) · [Fireworks AI bulletin](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) · [Notion advisory](https://www.notion.com/en-gb/security/advisories/security-incident-involving-a-third-party-ai-vendor)

### 3. Open source · llama.cpp and ggml reach v1.0.0 with a stable API

About 3.5 years after it started, the most widely used local-inference library now promises semantic versioning. Breaking changes to the public C API will now come with a major version bump.

- `libllama` (llama.cpp) and `libggml` (ggml) both ship v1.0.0.
  - MAJOR means a breaking change to `./include`, MINOR means backward-compatible features, backends or model support, and PATCH means bug fixes.
- Distribution now covers Docker, Debian, Homebrew, conda-forge, MacPorts, Winget, an XCFramework for iOS and macOS, and pre-release builds for Android and iOS.
- It supports 150+ architecture variants and all mainstream speculative-decoding methods.
  - CI runs on HF Buckets and HF Jobs, with self-hosted runners from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.
- Clément Delangue's take: "We're going to win by building, in the open."
  - The team also points to its Llama desktop apps for macOS and Windows (there's no Linux app yet, as one commenter noted).

**Why it matters:** Anyone building on llama.cpp (Ollama-style wrappers, on-device apps, your local experiments) can now pin a major version instead of chasing a moving `master`.

Sources: [HF blog — llama.cpp v1.0.0](https://huggingface.co/blog/ggml-org/llama-cpp-v1) · [@ggerganov](https://x.com/ggerganov/status/2121246103312212124) · [@ClementDelangue](https://x.com/ClementDelangue/status/2121260104913768931)

### 4. Assistants · Gemini can now read your spending through Google Wallet (and Plaid-linked banks)

Every US Gemini app user can connect Google Wallet to Personal Intelligence, and can link bank, card, loan and investment accounts through Plaid for more detail. Access is read-only: Gemini can't move, manage or invest money.

- Tag @Google Wallet and ask things like "what subscriptions went up in price this year?" or "set my holiday budget from last year's gifts and travel".
  - Gemini can answer with live charts of cash flow and budgets.
- It's rolling out on web, Android and iOS, and you can switch the connection off at any time under Connected Apps.
- The catch, per The Next Web's reading of Google's help pages: Google doesn't train directly on Wallet data.
  - But with Gemini activity on, it may train on summaries of your transactions, and human reviewers may read some chats.

**Why it matters:** OpenAI's agents can now spend and Google's can now see your money, both on the same day. Assistant competition has moved into personal finance, and each company is starting where it can earn trust (read-only at Google, approve-every-purchase at OpenAI).

Sources: [Google blog](https://blog.google/innovation-and-ai/products/gemini-app/google-wallet-gemini-app/) · [Plaid](https://plaid.com/blog/google-wallet-gemini-plaid/) · [The Next Web](https://thenextweb.com/news/google-gemini-wallet-spending-insights-plaid)

### 5. Robotics · OpenAI disbands its stand-alone robotics team and bets on general models running robots

Fortune reports, and OpenAI confirmed, that robotics researchers moved into main research under Mark Chen, while applied robotics and hardware went to Applied under Greg Brockman. OpenAI says this speeds robotics up rather than retreating from it, and the humanoid demo is still targeted for 2027.

- OpenAI says that at the start of the year it planned to build robotics on world-model work like Sora.
  - Since then, general frontier models, GPT-6 Astra included, have proved they can be applied to robots directly.
- One data point Fortune cites: in an NVIDIA study, an Astra-driven agent reached 74.8% on a simulated manipulation benchmark, against 51.0% for the best specialised robot models on that leaderboard.
- OpenAI's booth (#208) at CoRL in Austin was empty this week.
- Researchers pushed back. CMU's Guanya Shi said you still need small on-robot models that work without a network. Russ Tedrake (Walden Robotics) said real-world deployment data is still the moat.

**Why it matters:** This is the clearest bet yet by a top lab that robotics is a downstream use of general intelligence, not a separate problem. It questions the thesis behind billions in robot-brain startup funding.

Sources: [Fortune via Yahoo — OpenAI disbands stand-alone robotics team](https://tech.yahoo.com/ai/articles/exclusive-openai-disbands-stand-alone-170530559.html)

### 6. Robotics · NVIDIA GR00T 2: an open world-action model that learns skills from human video

GR00T 2 predicts future video and robot actions together. NVIDIA says it transfers skills between robot bodies without robot demonstrations, and five demos per task bring it to about 95% success. New demo videos went out on 13 November, and the weights are due in December.

- It's built on Cosmos and pretrained on:
  - 402,730 h of egocentric human video
  - 10,184 h of handheld UMI-gripper demos across 2,897 tasks
  - 23,500 h of teleoperation
  - Robots covered: YAM and Franka arms, Sharpa and Wuji hands, and the Unitree G1
- Results on 50 tasks the target robot had never seen:
  - Zero-shot: **40%**
  - With 5 demos (under 3 minutes of robot data plus a 30-minute fine-tune on one RTX 5090): **95%**
  - Two 22-DoF hands seat GB300 memory modules 85% of the time
- Agentic control: a reasoning model calls GR00T 2 through the WAM API and gets 97+ of 100 consecutive toy assemblies right.
- Open now: Human-1M-100K (100k hours of human video, licensed for commercial use) on Hugging Face, SimFoundry 1.0, and the UMI/DexUMI hardware designs.
  - The GR00T 2 and Turbo weights follow in December, but the license isn't published yet. NVIDIA's Flora models shipped under a non-commercial license.

> Jim Fan's line: "You need to own System 1, untethered and free of the tyranny of conference Wi-Fi."

**Why it matters:** It's the opposite bet to item 5. NVIDIA argues for a dedicated "System 1" model on the robot, steered by a big reasoning model.

Sources: [NVIDIA Research — GR00T 2](https://research.nvidia.com/labs/gear/gr00t-2/) · [NVIDIA blog (CoRL)](https://blogs.nvidia.com/blog/robotics-simulation-advancements-corl/) · [@nvidia](https://x.com/nvidia/status/2121319866712076769) · [Humanoids Daily](https://www.humanoidsdaily.com/news/nvidia-gr00t-2-world-action-model-corl-2026)

### 7. 3D · World Labs opens Atlas to everyone, with photo-to-world, Chisel and a developer API

Atlas, World Labs' "omni" model (text, images, video and 3D in one shared spatial context), moved from partner early access to self-serve on 12 November. It launched with tools to reconstruct, block out and export worlds as meshes. Public pricing hasn't been published.

- Viewfinder reconstructs a scene from photos, reportedly up to 64, or generates a world from a single image and renders camera paths.
  - Chisel lets you graybox a scene and see every camera view update. Splats to Mesh and Relight handle export and lighting.
- There's also a developer API (reported endpoints include `atlasGenerate`, `atlasChisel`, `images2PosedRGBD`, `splats2Mesh`) and a Blender integration.
- Architecture: a multimodal autoregressive diffusion transformer that takes camera poses and depth maps as native inputs.
  - It renders camera-controlled video up to 1 minute at 1440p and outputs point clouds or Gaussian splats.
- In World Labs' own camera-control test with human raters, Atlas was preferred 75% of the time against MiniMax H3 and 81% against Gemini Omni Flash.
  - World Labs notes that rivals received the camera motion as text prompts, which works in Atlas's favour.

**Why it matters:** "Photos in, editable 3D mesh out" through an API is a real workflow for architecture, product shots, VFX previz and robot simulation, without a photogrammetry pipeline.

Sources: [World Labs — Atlas](https://worldlabs.ai/atlas) · [Daily AI Primer — Atlas technical summary](https://www.ai-primer.com/engineer/stories/world-labs-atlas-world-model) · [Genativity — public access](https://genativity.com/news/world-labs-atlas-public-access-no-published-price)

### 8. Dev tools · Claude Code: a more proactive Projects coordinator, plus custom themes through mods

In Claude Code Projects, the coordinator agent now reports on its own in the project chat (status, decisions waiting on you, finished work, blockers). On desktop you can reskin Claude Code with themes built as mods.

- Projects: the coordinator answers questions in project chat and posts updates on parallel threads.
  - Reported extras include reminders, read-only thread sharing, `/btw`, `/fork` and a better threads panel.
- Themes: update to v2.1.324+, install `theme-examples@claude-plugins-official`, then run `/custom-theme`, or just describe the look you want. Themes can set colours and animated backgrounds.
- Security note: mods run with your permissions, so review a theme before installing it, especially after item 2.
- Elsewhere: Theo shipped in-app UI customisation for T3 Code the same night ("ask your agent to make changes and it will update on the fly").

**Why it matters:** The coordinator update matters more than the themes. It moves Projects closer to a manager that pings you only when it needs a decision, the same pattern as Cursor Projects.

Sources: [@ClaudeDevs — Projects](https://x.com/ClaudeDevs/status/2121375309287301360) · [@ClaudeDevs — themes](https://x.com/ClaudeDevs/status/2121392666244186327) · [Claude Code changelog](https://code.claude.com/docs/en/changelog) · [@theo](https://x.com/theo/status/2121455838414201200)

### 9. Enterprise · Claude for Manufacturing: 17 skills, 4 connectors, and Infor write access

Anthropic's new plugin for Team and Enterprise plans covers quoting, failed inspections, breakdown diagnosis, re-sourcing, design variants and production reports. It connects to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite, and every flow ends with a human review.

- Example flows: an estimator has Claude read a request, list what's missing, compare similar jobs and draft a quote.
  - A technician feeds in sensor exports, repair notes and photos plus the manual and gets ranked likely causes.
- Infor ships its own plugin, so Claude can make changes too, such as releasing a purchase order to another supplier and rescheduling production.
  - It acts within the buyer's own Infor permissions and after their approval.
- Systems without a connector work through regular exports saved to Microsoft 365.
- It doesn't operate equipment. For that, Anthropic points to its Model Hardware Standard research preview. A technical webinar is on 17 November at 11:00 ET (17:00 Warsaw).

**Why it matters:** This is the vertical-plugin playbook (skills plus connectors plus "Set us up") applied to plants, and it shows what a sellable industry bot package looks like.

Sources: [Claude — Introducing Claude for Manufacturing](https://claude.com/resources/articles/claude-for-manufacturing)

### 10. Safety · Anthropic's 10 rules for safety evals that actually mean something

An Anthropic Engineering post applies psychology's replication-crisis reforms to AI safety evals. If you only do three things: write down the construct before collecting data, validate your grader against humans, and report uncertainty.

- A safety eval asks "given this input, did the model do the right thing?". It measures propensity to cause harm and how severe that harm is, not capability.
- Known failure modes: evals are easy to game, often don't measure what they claim, are hard to reproduce, and give results that shift between multiple-choice and open-ended formats or even when answer order changes.
- The fixes borrow from psychology: preregistration, mixed teams of domain and methods experts, grader agreement measured with statistics like Cohen's kappa, running evals safely, and responsible disclosure when an eval finds a flaw in a deployed system.

**Why it matters:** Every "safer than its predecessor" claim in a system card rests on evals like these. The post gives you a checklist for judging those claims, and for checking your own bot evals.

Sources: [Anthropic — Building strong safety evaluations](https://www.anthropic.com/research/building-strong-safety-evaluations) · [@AnthropicAI](https://x.com/AnthropicAI/status/2121317227618340990)

## GitHub repos

- **[google/elephant-goldfish-model](https://github.com/google/elephant-goldfish-model)** (Apache-2.0, ★18.7k, +5.5k this week) — "Design with an elephant. Review with a goldfish." It's a plugin for Claude Code, Codex, Antigravity and ChatGPT desktop.
  - The Elephant agent holds your project context and drafts a spec. Goldfish reviewers then read the draft and the referenced sources *without* the earlier conversation, to catch unstated assumptions.
  - There's a 60-second review mode for an existing README or PR plan checked against live code.
- **[Tencent-Hunyuan/Hy-MT2](https://github.com/Tencent-Hunyuan/Hy-MT2)** (★1.75k, trending today) — "Fast-thinking" translation models in 1.8B, 7B and 30B-A3B MoE sizes, covering 33 languages.
  - With AngelSlim 1.25-bit quantisation the 1.8B model shrinks to 440 MB for on-device use. Tencent claims the 7B and 30B models beat DeepSeek-V4-Pro and Kimi K2.6 in fast mode.
  - The license isn't a standard SPDX one, so read it before commercial use.

## Research papers

- **SO-Arena** (2611.15777, top paper, 82 upvotes) — a benchmark of "System One" decision models across 11 domains and five capability families. It measures how reliable their decisions are, not only aggregate scores. [HF](https://huggingface.co/papers/2611.15777)
- **Actor, Judge, or Teacher?** (2611.17167) — can a decision model that acts badly still teach well? It compares three decision models and three LLMs on ALFWorld and WebShop, with JEV as the RL teacher. [HF](https://huggingface.co/papers/2611.17167)
- **PhysLoop** (2611.16748) — a harness that uses LLMs' physics knowledge to produce correction pairs, so video models can improve themselves on physical consistency. [HF](https://huggingface.co/papers/2611.16748)
- **Duo** (2611.15131) — a "Reciprocal World Model". It's full-duplex, perceiving audio and video continuously while it answers in real time. [HF](https://huggingface.co/papers/2611.15131)
- **Open Vision Banana** (2611.15844) — a fully open recipe for turning pretrained image generators into multi-task perception systems, an open replication of the closed Vision Banana. [HF](https://huggingface.co/papers/2611.15844)

## AI workflows

**1. [A free morning email brief across four mailboxes, with drafted replies](https://x.com/AbhayPerspctve/status/2121107718975230431)**
- **Replaces:** Going through four inboxes by hand every morning, sorting out promos, and writing routine replies for a role-based mailbox (an apartment association).
- **What to connect:** A Python script written by Claude, run daily on GitHub Actions and triggered by cron-job.org.
  - It uses the Gmail API (OAuth app published to Production, because refresh tokens in Testing mode expire after 7 days) and Gemini Flash on the free tier, with forwarding to bring non-Gmail mail in.
  - Output: one brief in five categories (Action Required, Transactional, Newsletters, Promotional, Social), promos trashed in one mailbox only, and reply drafts saved to Drafts.
  - It never sends a reply on its own. Running cost: ₹0 at the author's usage.
  - Free-tier Gemini terms allow human review of what you send, so test with made-up mail first.
- **Who it's for:** Non-coders and solo operators who juggle several personal and role mailboxes and don't want to pay for a no-code subscription for a job that runs once a day.

**2. [GPU inventory tracking by asking a bot instead of keeping a dashboard](https://x.com/ivanburazin/status/2120897158261952962)**
- **Replaces:** A custom vibe-coded dashboard that pulled together live capacity, contracts signed but not yet delivered, and contracts still in negotiation.
- **What to connect:** Grok Bot with Slack access. You ask "what's the current state of our GPUs?" and it searches the relevant Slack messages and threads and writes up the answer.
  - The author admits it's "not very efficient in the traditional sense" but says it works very well.
  - A reply in the thread adds a sensible check: spot-test a few answers against the contract records.
- **Who it's for:** Small ops or sales teams whose real status lives in chat threads, not a database, and whose questions change too often to keep a dashboard up to date.

**3. [Faster bot page updates: edit only what changed](https://x.com/VampyreDrakul/status/2121261805771038956)**
- **Replaces:** Having a Grok Bot rewrite and re-read a whole large page for every small change, which took about 5 minutes per update.
- **What to connect:** No new tools, just a change in how you ask. Request targeted edits so the bot changes and re-checks only the edited part.
  - The author reports 5 edits in about 47 seconds, down from about 5 minutes for one update.
- **Who it's for:** Anyone who keeps long living documents, trackers or pages maintained by an AI assistant and wants quicker, cheaper updates.

## Quick hits

- **The FAA ground stop wasn't Microsoft or AI** — The nationwide halt on 13 November lasted more than four hours and was the longest in 25 years.
  - Officials blamed ageing FAA telecom switches overloaded by new equipment, and ruled out both the Microsoft 365 outage that began the day before and SMART, the FAA's new AI traffic-planning tool.
  - A cyberattack hasn't been fully ruled out, but there's "zero evidence" of one. [GeekWire](https://www.geekwire.com/2026/dhs-chief-sees-no-microsoft-link-to-faa-ground-stop-says-possible-ai-role-being-investigated/)
- **Tesla opens Cybercab orders for personal use** — Select customers can order one for $29,990, with a mandatory $299/month Autonomy subscription (first month included) and a $1,000 non-refundable deposit.
  - Deliveries are estimated for December in Austin, and the car only works inside supported service areas. [Electrek](https://electrek.co/2026/11/13/tesla-cybercab-personal-use-orders-299-subscription-geofence/)
- **Google tested its agent anomaly detector on benchmarks it didn't write** — Across 1,774 sessions on AgentDojo, InjecAgent, ASB, AgentDyn and AgentHarm:
  - It flagged **87%** of harmful instructions that were carried out, with false alarms on 2–16% of benign sessions.
  - Its weak spot is poisoned long-term memory: it caught only **4 of 18** cases. [Google Developers Blog](https://developers.googleblog.com/how-agent-anomaly-detection-scores-on-benchmarks-we-didnt-write/)
- **Lebrel gives away an uncensored DeepSeek V4.1 Flash for four hours a day** — During its open beta, `lebrel/deepseek-v4.1-flash-uncensored` is free daily from 15:00 to 19:00 UTC (16:00–20:00 Warsaw) in chat.
  - It is also available through an OpenAI-compatible, attested API at `attested.lebrel.ai/v1`.
  - Outside that window it costs credits. [Lebrel](https://lebrel.ai/) · [@LebrelAI](https://x.com/LebrelAI/status/2121287151966900715)
- **"Stand with Jensen" is growing** — An independent open letter asks Nvidia's board to reinstate Huang while an independent investigation runs.
  - The Times of India reports more than 40,000 signatures, Elon Musk's among them, and signers on the timeline include @petergyang and @TheAhmadOsman.
  - The campaign says Nvidia and Huang don't endorse it. Nvidia's Q3 results are still due on 17 November. [Times of India](https://timesofindia.indiatimes.com/technology/tech-news/elon-musk-and-over-40000-others-including-nvidias-current-and-former-employees-sign-open-letter-for-jensen-huangs-return-letter-says-thirty-three-years-of-leadership-do-not-place-anyone-above-/articleshow/135823053.cms) · [NVIDIA](https://nvidianews.nvidia.com/news/nvidia-announces-ceo-transition)
- **Amodei, speaking with Jamie Dimon ahead of the IPO,** said Anthropic's hardest business problem is forecasting compute.
  - It buys conservatively up front and uses the spot market when demand beats its plan, and it aims to "underpromise and overdeliver". [Business Insider](https://www.businessinsider.com/anthropic-ipo-pitch-dario-amodei-compute-dilemma-2026-11)

## Go deeper

- [Cursor's incident update](https://cursor.com/blog/security-incident-update): read it as a template for what to check with your own AI vendors.
- [Anthropic — Building strong safety evaluations](https://www.anthropic.com/research/building-strong-safety-evaluations): the full 10 recommendations with system-card examples.
- [DetLLM: eliminating nondeterminism with integer-only arithmetic](https://nathan.rs/posts/detllm/): run every part of an LLM (matmuls, softmax, norms, RoPE) with no floats, and get bit-identical outputs on any GPU or CPU.
- [OpenAI Commerce Partner Playbook (via the Wallet post)](https://openai.com/index/introducing-chatgpt-wallet/): what a shop needs so agent purchases aren't declined.
