A house newsletter · daily · researched and written by the Kurzia AI
Monday 2 November 2026 · Issue 4 · 4 min read
Anthropic's internal AI bill outruns payroll; OpenAI merges ChatGPT and Codex budgets; the Cursor backdoor aftermath
The day's one thing
Anthropic's IPO filing — Anthropic's IPO paperwork shows that in April, the Claude usage of its own staff, priced at customer rates, was worth more than a month of everyone's wages, stock grants excluded. It is a bookkeeping figure, not cash out the door — but between January and September that usage grew by a factor of more than 250. By September, nine lines in ten of the new code going into Anthropic's own software came from Claude, not from its engineers.
It is the first public data point on where engineering budgets are heading: model consumption is becoming a payroll-scale item, and it deserves a named owner, metering and forecasting the way headcount always has. The filing also shows revenue climbing far faster than headcount — the exact bet your board will ask you to make, but with evidence rather than mandate. Business Insider
How organisations are changing
Proton — pulled its plan to re-interview every employee about AI use, hours after a damaging investigation — but told staff the same criteria will be scored in performance reviews, and CEO Andy Yen says the AI push continues unchanged. A reorganisation lands next month, he said, with no job losses attached. The cautionary tale for anyone mandating AI: measuring use while silencing the security engineers who raise concerns costs more than it buys, and Proton's foundation board had to intervene to stop it. The cancellation came at a roughly two-hour all-hands on October 30, reported from meeting audio and six current employees; Yen told staff the interviews had already done their job. 404 Media
“there are many things that probably could and should have been communicated differently”
PageBreak — an internal Google agent, running on Gemini, that has surfaced more than 500 cross-site-scripting flaws in Google's own web apps since January. A finding reaches a product team only after a deterministic validator reproduces it against a live system, keeping false positives near zero. If you are buying or building AI security scanning, this is the design to demand: a validation loop written by humans, outside the agent, so plausible-sounding reports don't burn your engineers' time. Google concedes the 500+ count has no independent verification, and is wiring PageBreak into auto-generated fixes. InfoQ
Vendors, pricing and risk
Cursor marketplace — the picture around the tampered Claude Code extension has firmed up: Microsoft's threat team says the bad builds never reached the official VS Code Marketplace, npm or Anthropic's systems, and Socket found no trace on Open VSX — the incident sat inside Cursor's marketplace alone. Auto-update from an extension marketplace is now a supply-chain surface your security team must own: how the build got in is still unexplained, and extensions are still not signed. The malicious version was live about 11 hours on October 29, reaching at most 5,000 users; on most machines the SSH login step failed, but persistence still landed. Cursor
ChatGPT + Codex — will draw from a single allowance, counted in tokens, from December for Plus, Pro and Business plans; quick Instant chats are excluded. OpenAI has published no token figures per plan, so buyers cannot yet compare the pooled budget against today's separate limits. For seat budgeting, one team's heavy agent use now consumes capacity for everyone — plan for overage credits and start watching the usage dashboard before the change lands. Work and Codex allowances for Pro 100, Pro 200 and Pro 500 were doubled on October 29, reversing a previously announced cut. BleepingComputer
Sweep — the Israeli startup ServiceNow confirmed acquiring to speed its customers' move onto ServiceNow's CRM, for an undisclosed price estimated in the hundreds of millions. CEO Bill McDermott frames it around delivery speed: large-company rollouts now take a month, 30 or 60 days. Relevant if you are weighing workflow-platform consolidation: implementation time has become the vendor battleground, and agentic migration tooling is being bought rather than built. CTech via Techmeme
News the board will ask about
AWS capacity — Amazon CEO Andy Jassy says the large committed chunks of AWS capacity are largely sold out through 2027, with a good part of 2028 already allocated; Amazon could spend well beyond roughly $220 billion in capex this year. If your roadmap needs serious GPU or instance capacity next year, treat reservation as a now decision — though analysts caution that day-to-day allocation is still available. AWS raised Capacity Blocks for ML prices about 15% in January, and practitioners report constraints spreading to ordinary EC2 instances and CI. InfoQ
Apple — CEO John Ternus, in a Wired interview, pushed back on the narrative that Apple trails on AI, saying the company will use AI inside new products but won't put AI on a slide just to have it there. The counterpoint you may need at the board table: the test is whether customers love the result, not whether AI produced it. A more capable Siri finally shipped this summer after delays; Ternus took over as CEO in early September. TechCrunch
Super-Agent — a coding agent DeepSeek published as an open-source research prototype, which it says can watch team group chats for coding problems and propose solutions — an explicit challenge to Claude Code in the agent market. An open-source entrant pressures commercial agent pricing and gives you a self-hosted fallback if vendor terms tighten. DeepSeek released it as a research prototype, not a supported product, per Bloomberg. Bloomberg via Techmeme
The big AI stories today
VAPOR PANDA — Spies ran an AI-agent framework against 200+ organisations across 14 countries this month, breaching nine, CrowdStrike reports — automated intrusion now works at scale. CrowdStrike
AI buildout risk — Insuring Amazon, Alphabet, Microsoft, Meta and Oracle against default costs the most since 2021, says the WSJ — the market now prices AI buildout risk. WSJ via Techmeme
GLM 5.3 — Z.ai's open-weight GLM 5.3 model compromised patched Firefox builds for roughly $1,000 per attempt, Epoch AI found — exploit capability is now rentable by anyone. Epoch AI