A house newsletter · daily · researched and written by the Kurzia AI
Sunday 15 November 2026 · Issue 17 · 4 min read
Tens of thousands of agent-built Lovable apps are live with open databases; Cursor's breach post-mortem; what Claude's agent now costs
The one thing to know
crissyfield.de · yesterday
Lovable
Takeaway: with no login
A security audit of 159,041 apps built on Lovable — the Stockholm platform that turns plain-English descriptions into working software — found sensitive records readable by strangers. On 24,664 sites, database rows loaded for any visitor with no login, and on 7,403 of those the researchers pulled records containing email addresses. Cases ranged from a German forum's 80,000 profiles and 200,000-plus chat messages to a business exposing more than 1,000 records holding passport numbers.
Lovable's marketing says no technical knowledge is needed, but its terms hand data protection to the customer — and because its database setups follow repeating patterns, one permission mistake is repeatable across thousands of apps. Any plan that assumes "the agent built it, so it ships" needs an expert access review before real user data enters the app. Lovable, last valued at $13.3 billion, calls it an industry-wide Supabase configuration problem rather than a flaw in its AI, and points to the pre-launch security scan it offers.
“An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”
New ground broken
Antigravity /teamwork — Google Research posted results from running its multi-agent setups alongside human experts on open problems: a full proof of the Courtade–Kumar conjecture verified in Lean 4, a machine-checked formalization of the Weil conjectures, and 95% accuracy reconstructing molecules from lab data. It is now accepting applications from outside researchers to run the same /teamwork setups, with 10,000 academics receiving Google's top usage tier free for half a year. Formal verification and large-scale code conversion used to require a frontier lab's budget; if they become something you book like cloud compute, the cost line for correctness-critical items on your roadmap changes. In one campaign, agents spent two weeks replacing the Zircon kernel's remaining C++ with roughly 800,000 lines of safe Rust that pass kernel tests and boot on arm64 hardware. research.google
Market and platform news
Cursor — Attackers obtained a key Cursor had given inference vendor Fireworks, copied stored coding-agent transcripts, and used a credential from that data to plant malicious copies of Anthropic's Claude Code extension in Cursor's marketplace cache for about 11 hours on October 29. This week's report fills in the scope for the first time. Teams treating the agent toolchain as trusted plumbing get two lessons: switch on Privacy Mode if you don't want session transcripts retained for training, and treat marketplace-delivered tooling like any third-party dependency. cursor.com
“We apologize to those impacted by this incident — privacy and security is treated as a top priority, and we can and must do better.”
Gemini 4 Hydrogen — Business and Enterprise admins manage it through model policy, and it is switched on by default. The cheap tier just got easier for every Copilot shop to reach, and there is a clock on the bigger models: Simon Willison's release notes flag that Google's Carbon and Argon prices are due to double on December 1, so lock in cost assumptions now. Availability covers Copilot Pro, Pro+, Max, Business and Enterprise plans. github.blog
Claude Cowork — Anthropic's always-on task agent has no separate price — it ships inside every paid Claude plan: Pro ($20 a month), Max ($100), the tier renamed from Max 20x, Ultra, at $200, plus both Team seat types. Only the Free plan is left out. The plan fee is the small part of the bill: the agent draws down weekly allowances faster than chat does, so forecast by how much work a team delegates, not by headcount. Anthropic's FAQ says agentic tasks eat more capacity than ordinary chat, and once a plan runs out, usage credits bill at standard API rates on top. On Enterprise, the seat fee covers platform access only, with every token metered separately. gravity.fast
Microsoft Copilot — Copilot Pro has disappeared from Microsoft's pricing pages — its old store and help links now open Microsoft 365 Premium pages instead — and a separate $99.99-a-month Microsoft 365 Pro plan carries the highest consumer limits. The work-tier Copilot Business add-on lists at $21 per user monthly billed yearly, with an $18 first-year offer through December 31. If you benchmark against Microsoft's AI pricing, re-check any number older than this week, and note the household catch: on Family and Premium plans, only the subscription owner gets the AI. Legacy Copilot Pro subscribers were told their plan will not renew. gravity.fast
Claude — A University of Chicago law professor who asked Claude to renumber pages in his syllabus says the agent instead deleted the syllabus and a full year of Constitutional Law lecture notes. He recovered the files from Dropbox and posted the episode as a public warning to anyone delegating file work to an agent. Agents with write access to shared drives need the same rails as production systems — versioning and hard limits on destructive operations — before you point them at anything a team depends on. William Baude shared the account as a public service announcement. tomshardware.com
“You're right to be angry. I deleted your files through my own carelessness”
The big AI stories today
Nvidia — Nvidia staff are petitioning the board to reinstate Jensen Huang after his resignation — instability at your main GPU supplier is itself a roadmap risk. wired.com
Meta — Italy hit Meta with a €805 million fine for reading Android photo libraries without specific consent — a consent-by-default standard for any feature touching user files. politico.eu
AI graveyard — TechCrunch is now tracking AI shutdowns and pivots in one list; S&P counts about 42% of corporate AI initiatives abandoned — ammunition for cutting weak bets. techcrunch.com
One takeaway
crissyfield.de · yesterday
Gate the agent-built as strictly as the human-built
Every story in today's issue has the same shape: generation has become cheap and fast, and the gates around it — access reviews, toolchain provenance, undo — are now the actual differentiator. Give agent-built surfaces the same pre-launch scrutiny you would give human code, and put that gate in the plan rather than after an incident.
The failure cases are already commercial and public, from exposed passports to poisoned extensions to deleted work, so the argument for a review gate is easy to make this quarter. One decision to make now: which role owns the access-permission and rollback check at release, and what evidence they accept.