Amazon's 98% model swap, Notion's vendor breach, and Anthropic's $1.4T IPO range

🔥The day's one thing

🏢How organisations are changing

• Cubist — Point72's quant unit Cubist claims in a lawsuit that a fired lead research analyst used Claude and OpenAI's Codex to copy source code before joining a competitor. Agent logs are becoming legal evidence; expect them to be subpoenaed, and check what your own coding-agent tooling records and how long it retains it. It is an allegation in a filing, not a finding — but the named tools are the coding agents your engineers already use. Techmeme (via Bloomberg)

• Lenny's tech job market review — An analysis of the late-2026 tech job market argues product managers are gaining most in the AI era, engineer demand is plateauing, designers are losing out, and the biggest AI hiring boom sits outside software. If you are hiring engineers into a cooling market, budget leverage is shifting to you — and this is the framing the board will bring to any headcount conversation. It is a newsletter's synthesis rather than a study; treat it as a talking point, not evidence. Lenny's Newsletter via Techmeme

• AI usage generates a fragile codebase — A widely shared essay by an engineer laid off from Meta argues heavy AI coding makes codebases fragile: shallow tests, over-complex code, shrinking human understanding, and pull requests arriving faster than review can absorb. His warning signs — incidents creeping upward, recovery dragging out — are ones you can check against your own metrics this quarter, whatever you think of the argument. He says 90% of his code was AI-generated in his final months at Meta, the floor for a decent rating; it is one engineer's account, not a measured study. evanxmerz.com

  “AI codebases are software made by AI for AI.”

🧾Vendors, pricing and risk

• Fireworks AI breach — Notion says a breach at Fireworks AI, the vendor it used to classify Notion AI chats and serve open-weight models, may have exposed chat transcripts tied to fewer than 1% of workspaces. Your AI assistant's subprocessors inherit your data — ask every vendor who actually serves the model, and what happens to chats sent there. Notion cut Fireworks off on November 3 and deleted the data; it advises rotating any credentials anyone pasted into Notion AI. Notion Help Center

• Grok Build — Security firm Claroc found that Grok Build's shareable app links put their random part in the subdomain, where passive DNS search collects them: over 35,000 of 43,601 links were open to anyone. Business teams building internal tools on vibe-coding platforms can publish customer and staff data without engineering ever knowing — a blind spot for your data-loss tooling. Exposed material included a hospital roster of about 3,565 employees and 24 SpaceXSI internal apps, among them a sales map of 594 accounts. Ars Technica

• Git 2.56.1 — Git shipped fixes for six security flaws, among them attacker-controlled code execution during a recursive clone of a hostile repository and heap corruption during big merges. Agents clone and merge untrusted repositories at machine speed — patch Git everywhere it runs, including CI hosts and any tool that bundles its own copy. Patched builds are out for every supported GitHub Enterprise Server line back to 3.19, and GitHub.com itself is already fixed. The GitHub Blog

• RAISE Act — New York opened registration for the RAISE Act, its frontier-model safety law: developers above $500 million revenue must publish safety frameworks, report critical incidents within 72 hours and file transparency reports, registering by January 1, 2027. Three states now run comparable regimes (New York, California, Illinois) — a procurement-diligence item for model vendors, and a 72-hour clock that may flow into your contracts. Penalties reach $1 million per violation and $3 million for repeats, under a new state office, DIGIT. TechCrunch

📰News the board will ask about

• Anthropic IPO — Anthropic set an IPO range of $70–$78 a share for 825 million Class A shares and started the roadshow; Nasdaq trading as ANT is expected close to November 19. The supplier behind much of your agent stack will answer to public markets carrying a nine-month net loss of $104 billion on $33.2 billion of revenue — watch enterprise pricing and contract terms after listing. A $26 billion private placement of non-voting stock goes to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC, with Google and Nvidia putting in $10 billion each. TechCrunch

• Claude Meetings — Claude can now join video calls as a visible participant and draft plans, timelines and answers while people talk; collaborative Whiteboards and Sheets arrived alongside, and Claude Data left beta. An AI assistant sitting in internal meetings is a data-governance question before it is a productivity one — establish where those transcripts land before teams enable it. It is in beta on Team and Enterprise plans with Google Meet only, and Claude leaves automatically if an outside guest joins. TechCrunch

• ChatGPT Business Agents — OpenAI launched verified Business Agents that represent companies inside ChatGPT, plus a Business Hub where firms claim a profile and see mention counts, website clicks and discovery topics. If customers ask ChatGPT about your products, your comms and e-commerce teams now have a channel to claim — and a reason to know what it already says about you. The beta is limited to eligible US e-commerce and online service businesses, and OpenAI is not saying how verification works. The Verge

🌍The big AI stories today

• AlphaProof Code — Google DeepMind says AI agents formally verified Signal's actual code in four weeks and is open-sourcing the proofs — deep verification is becoming an agent task. Google DeepMind

• CoreWeave Q3 — CoreWeave grew revenue 168% to $3.66B but lost $709M, with backlog at $132.5B — the compute under your AI vendors is still bought years ahead. CNBC via Techmeme

• Railway hack report — Railway says its October breach targeted AI provider keys for resale, with about $740K of unauthorised spend — cap and rotate the keys your agents hold. BleepingComputer via Techmeme

💡One takeaway