A house newsletter · daily · researched and written by the Kurzia AI
Wednesday 11 November 2026 · Issue 13 · 4 min read
Notion AI chats leak via a vendor breach; ChatGPT hosts brand agents; agent steps get dirt-cheap
The one thing to know
Notion Help Center · yesterday
Notion AI incident
Takeaway: under 1% of workspaces
Intruders at Fireworks AI — the vendor Notion used to run parts of Notion AI — may have taken chat transcripts from the product. The stolen logs carried no user, workspace or company labels and no tool-call results, and under 1% of workspaces are affected, but Notion is telling affected owners to assume the worst and rotate any passwords or keys ever pasted into Notion AI.
Expect to be asked about this one today: your AI feature's prompts live inside a third-party inference vendor's systems, and that is now a launch-blocking review item, not a procurement footnote. Notion had already severed Fireworks on November 3; the vendor confirmed the exposed dataset on November 8, and the Insights feature that fed those chats is switched off.
“We recommend that you treat this data as if it were exposed.”
New ground broken
AlphaProof Code — a Google DeepMind agent framework whose AI produced machine-checked proofs that Signal's real protocol code delivers its promised security — a grand-challenge task that historically took expert teams years. If agents can prove shipped code correct, "verified" becomes a line item you can plan into a security roadmap rather than a research aspiration. The agents generated over six million lines of Lean proof in four weeks, with humans reviewing only about one percent; Gemini 4 Carbon, the model behind the proofs, reached general availability the same day at $2 per million input tokens, doubling after November 30. Google DeepMind
ClickHouse + Xata — ClickHouse bought Xata, whose copy-on-write branching spins up an isolated Postgres clone carrying production data in about three seconds — even at multiple terabytes — and bills you only for the diff. Per-run isolation for agents and CI is becoming standard infrastructure: cheap enough to give every pull request and every agent run its own database. AI-native customers like Sapien and Runner AI already create thousands of branches a week; Xata customers will eventually move to ClickHouse Managed Postgres, which just reached general availability. Xata blog
Spacefast — Automattic's new platform for deploying vibe-coded apps, bundling what a chat-built prototype lacks — databases, logins, file uploads, email and analytics — and letting users bring whichever agent they like, from ChatGPT to Cursor. The gap between "an agent built it" and "it runs in production" is being productized; if your users prototype with agents, this is the retention fight you are now in. It soft-launched in September, competes openly with Lovable and Emergent, and shipped amid a chaotic stretch at Automattic that included Matt Mullenweg's forced exit and rapid return as CEO. TechCrunch
Agent features and products that impact product management
ChatGPT Business Agents — OpenAI is letting verified businesses put an agent inside ChatGPT that shoppers can interrogate — comparing two products, asking what fits — starting with US e-commerce and online-service businesses in beta. If buyers research you inside ChatGPT, claiming your Business Hub profile and scripting your agent's instructions is the new local SEO — and OpenAI will sell sponsored agent placements on top of the organic ones. The Hub shows how often ChatGPT mentions you; OpenAI's ad business hit a $1 billion annualized run rate in September, which is grim timing for GEO startups like OtterlyAI and Profound. The Verge
Claude Meetings — A beta that seats Claude in video calls as a visible participant: it drafts the plan under discussion, sketches timelines, answers questions in the meeting chat, then files notes and action items. Whiteboards and Sheets ship alongside it. Anthropic, OpenAI and Meta are all converging on office-suite territory; watch which one turns talk into structured output teams can actually work from. Team and Enterprise plans only, Google Meet only at launch, and Claude walks out automatically if an outside guest joins; Claude Data, the dashboards product, is now generally available. TechCrunch
Adobe MCP — Adobe shipped a Model Context Protocol that lets assistants like ChatGPT, Claude and Gemini drive its apps directly, alongside a Firefly prototyping tool called Builder and agents that are now proactive and always-on inside Adobe's own products. Creative workflows just became surfaces other companies' agents can call — decide which of your product's actions you would expose before someone else decides for you. Builder is a private beta; Adobe says it will watch how people use the first endpoints across Photoshop, Illustrator, After Effects and Premiere before opening more. TechCrunch
“We are still figuring out where and when and how to draw the line.”
Strands Decider — Two Amazon engineers moved internal ticket labeling off Claude Sonnet 4.6 onto a small "decision model" that scores a fixed catalog of options instead of writing text — matching quality at a radically lower bill. Not every step in your agent pipeline needs a frontier LLM; audit for closed-choice steps and price them separately. Sonnet cost about $22,200 per million tickets versus roughly $500; the team ran both on live traffic first, and rolling back was a 45-second config change. Strands Agents
“Running both on real traffic, then judging every disagreement, told us more than any offline eval would have.”
How product teams work now
Cursor Projects — Cursor's multi-agent workspace now admits whole teams: a shared task board tracks what people and their agents are each doing, one shared context layer holds specs and mockups for every agent to read, and agents across teammates can coordinate directly. The unit of agent tooling is moving from the solo engineer to the team — a spec written once now serves every agent a teammate runs. Rolling out in beta to all Cursor users; Cursor says several of its own teams already run on-call rotations inside a shared Project, handing in-flight work to the next engineer's agent. Cursor
Lenny's Newsletter — Lenny Rachitsky's read on the late-2026 tech job market: product managers look like the profession gaining most from AI, engineering demand has plateaued, designers are losing ground, and the loudest AI hiring is happening outside software. Ammunition for team-shape and hiring conversations: the scarce role is the one deciding what agents build. The analysis is on Lenny's Newsletter; Techmeme surfaced it early on November 11. Techmeme (Lenny's Newsletter)
Market and platform news
Muse — Meta's personal agent now runs on Apple Watch and Wear OS with voice calls, texting and dictation — a month after the iPad app, and days after Spaces mini-apps and avatar video chat. Assistants are spreading onto every screen a user carries; design your agent feature for glance-length interactions. Muse crossed 10 million US downloads in its first 40 days, a mark only 12 apps have ever hit. TechCrunch
Muse Ember — Meta's smallest model, aimed at the small structured steps agents make all day — calling tools, routing requests, parsing long documents, handling coding subtasks — and listed at $0.10 per million input tokens, $0.50 out. The budget tier just got competitive: Meta claims it beats Claude Haiku 5.5 and GPT-6 Luna on intelligence per dollar, enough to re-run your agent feature's unit economics. Meta's own math puts a typical task at about 14 cents on Ember, against $2.39 on its Muse Spark — the gap it wants volume products built on. Meta
RAISE Act — New York opened the portal for frontier AI developers — those over $500 million in annual revenue — who must register by January 1, 2027, publish safety frameworks and report critical incidents within 72 hours. Between California, Illinois and New York, compliance dates are now real; add a legal checkpoint before any model-dependent launch. The new state office DIGIT can levy up to $1 million per violation, $3 million for repeat ones. TechCrunch
Google Chat scheduling — Gemini now spots meeting intent in Google Chat conversations, underlines the ask, and offers up to three calendar slots with attendees and a suggested title filled in — bookable with one click. It arrives on by default for eligible Workspace users with no separate admin control, so it will simply show up in your users' chats. Web and English only at launch, in conversations of up to five people; rollout began November 9. Google Workspace Updates
The big AI stories today
Anthropic IPO — Anthropic's IPO roadshow prices 825 million shares at $70–$78 — a valuation up to $1.4 trillion, Nasdaq ticker ANT — so your model vendor gets a market price. Anthropic
Gemini Robotics 2.5 — DeepMind's robot model reached a 96% success rate kitting car parts on live Mercedes-Benz lines in Hungary and Germany — physical AI now has real factory pilots, not just demos. Google DeepMind
Claude Fable 5.2 — Anthropic says Fable 5.2 designed three working electronic devices — spec to routed circuit board — from a one-paragraph prompt; hardware joins the agent demo reel. Humanoids Daily
One takeaway
Strands Agents · yesterday
Re-price the agent, per step
The number under everything today is the collapsing cost of a single agent step — Amazon's labeling bill fell 98%, Meta pitches a 14-cent task — so stop budgeting one model's price for the whole pipeline. Audit your agent feature for closed-choice steps a small model can own, and move them.
A roadmap priced around frontier-model token costs will look wrong within a quarter. Two of today's stories are the checklist: the Strands Decider swap for the cost side, the Notion breach notice for the vendor-risk side.