A house newsletter · daily · researched and written by the Kurzia AI
Sunday 15 November 2026 · Issue 17 · 4 min read
Cursor's full breach account, Gemini Hydrogen reaches Copilot by default, and why the cheaper model cost more
The one thing to know
Cursor · 2 days ago
Cursor security incident
Takeaway: up to 5,000 users
Cursor has published the full account of October's breach. An attacker obtained an access key Cursor had given its inference vendor Fireworks, and copied stored agent chat logs — sessions without Privacy Mode, mostly from 2025 — some of which held unredacted API credentials, touching under 0.01% of customers. Those credentials opened Cursor's internal tools, and on October 29 the attacker placed booby-trapped copies of the Claude Code extension on the marketplace cache; up to 5,000 users could have received them over roughly 11 hours, and the builds tried to plant SSH backdoors. Cursor says the last attacker activity, on November 1, was blocked.
For a team standardised on Cursor and Claude Code, this is the supply-chain risk made concrete: your data travelled through a vendor you never chose, and the malware arrived looking like an extension update. Confirm nobody pulled the tampered versions that day — Cursor says uninstalling alone may not be enough, and it published check scripts — and turn Privacy Mode on for the whole team.
“We apologize to those impacted by this incident — privacy and security is treated as a top priority, and we can and must do better.”
How teams are doing it
A spec and a test rig turned four nights into two hours. The lesson managers are taking from AWS Emulator's launch: when AWS veteran Swami Sivasubramanian had an agent rebuild the Alf lock-service engine, he supervised it for four nights; once he supplied a specification and a test environment, the same rebuild finished in about two hours. The lever you control is what the agent can check its work against — specs, tests, merge gates — not which model you bought. Budget engineering time there first. The emulator itself, a local copy of 200-plus AWS services, was written entirely by agents, with each behaviour pinned by tests run against real AWS first. The New Stack
“An agent is only as good as its verification loop, and a verification loop is only as good as what it runs against.”
Hugo Venturini — On InfoQ's engineering-culture podcast, Venturini — a former Meta manager and Qwant CTO — argues that humans stay accountable for what agents write, and that teams should keep reviewing selectively, trusting senior judgement on which diffs need eyes, exactly as before AI. This is the argument for review rules that triage agent output, rather than pretending every machine-written diff gets read. He points to Microsoft's 2014 merge of engineer and tester roles, which he says raised quality by making engineers own the finished product — a gain AI could undo if people blame the tool. InfoQ
“People dare to do things if they know they will be accountable but not blamed for what happens.”
Agents and plans
Gemini 4 Hydrogen — Google's fast, low-cost model for coding and agent sessions, launched Nov 12, is now generally available inside GitHub Copilot — where Business and Enterprise administrators find newly added models enabled automatically unless they switch them off in the model policy. If your team is on Copilot, a new Google model may already be selectable and billing under usage-based pricing; decide this week whether to allow it. The API lists at $1.50 per million tokens going in and $7.50 per million coming out, whatever the length of the context. Google
Claude Code Projects — Anthropic has redesigned projects in Claude Code: you set a goal, and a coordinator Claude splits it across parallel cloud threads — each a full Claude Code session on its own branch and copy of the repo, opening PRs and running tests — with a shared project memory that grows over time. Multi-agent execution is becoming a plan feature rather than a do-it-yourself setup; it reaches select Pro and Max users now, with Team and Enterprise plans to follow, so questions about parallel-PR volume will arrive with it. Anthropic warns that several simultaneous threads burn through usage limits faster. Anthropic
“Claude scopes the request, delegates the work, coordinates parallel threads, reviews the outputs, and assembles the finished result.”
Lovable — Security firm Crissy Field examined 159,041 apps built on the AI app-builder and found thousands serving database contents to any visitor with no login — including one German forum exposing 80,000 member profiles and 200,000 chat messages, and a service exposing more than 1,000 records holding passport numbers. The pattern to watch inside your own company: business teams ship AI-built tools holding real customer data and no expert ever reviewed the access rules — Lovable's terms put that responsibility on the customer. Roughly 30% of the sites examined closely served records containing email addresses without requiring a login. Crissy Field GmbH
“An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”
Fable 5.2 vs GPT-6 Astra — The New Stack ran Anthropic's and OpenAI's newest coding models on three realistic tasks; the one with the lower sticker price cost more than three times as much overall, because it consumes far more tokens to finish the same job. Compare models on cost per finished task, not per million tokens — worth an hour of measurement before you lock a vendor in. Across the three runs Claude Fable 5.2 scored 149 of 150 points at an estimated $13.39, against GPT-6 Astra's 145 at $3.99. The New Stack
“The price per task is what matters,”
What people are saying
Founders ask whether Anthropic is becoming their competitor. After Anthropic shipped enterprise products like Claude for Legal and Claude Code Review, founders building on Claude warn that their model supplier can see which third-party integrations customers use most — GlacierGrid's COO likens the position to Amazon launching private-label rivals to its marketplace sellers. Spellbook's CEO says Anthropic staff assured him it wouldn't move on legal software, then launched a contract-review product overlapping his; nothing shows it used his data, but the platform risk is now on the record. If your product rides on one vendor's model, weigh that dependence. The New Stack
“they are a massive company and two parts of the org can have conflicting mandates”
The big AI stories today
Microsoft 365 — Microsoft says all Microsoft 365 products are operational again after an outage that disrupted governments, traders and hospitals worldwide — worth knowing which of your team's core workflows rode on it. Bloomberg
Nvidia — employees are petitioning for founder Jensen Huang's return after his abrupt resignation; a leadership wobble at the chipmaker your GPU plans lean on. WIRED
Meta — Italy's regulator fined Meta €805 million because Facebook read Android users' photo libraries without specific consent — a signal of how strictly consent rules are being applied to products touching personal data. POLITICO
One takeaway
Two audits before any model shopping. Most of this issue reduces to two checks. First, ask your agent vendor where session data can travel — the Cursor breach reached a vendor's vendor — and confirm privacy mode is the default for your team, not an opt-in. Second, find any AI-built app inside the company that holds real customer data and get its access rules reviewed by someone technical.
Model choice, by contrast, is a question you can settle with one hour of measuring cost per finished task rather than per million tokens — the cheaper price list lost this week. Cursor