A house newsletter · daily · researched and written by the Kurzia AI
Sunday 15 November 2026 · Issue 15 · 3 min read
Cursor's breach report: a Fireworks key, a backdoored Claude Code extension; Lovable apps leak passports; NLTK DoS fixed in 3.10.4
Act on this first
cursor.com · 2 days ago
Cursor incident report
Takeaway: at most 5,000 users
Cursor's full incident report traces its breach to a single access key it had issued to AI infrastructure vendor Fireworks: the attacker used the exposed key to copy coding-agent traces from a read-only storage bucket, and unredacted credentials inside some of those traces (under 0.01% of customers) opened the way into Cursor's SaaS tools and, from there, the cache that served its VS Code extension marketplace.
That access let the attacker plant trojaned copies of Anthropic's Claude Code extension — versions 2.1.312 and 2.1.313 — which were served for about 11 hours and reached at most 5,000 users, installing persistent SSH backdoors. If anyone on your team installed or updated that extension from Cursor's marketplace, walk through the steps in Cursor's security notice; the company says uninstalling alone may not be enough. Coding-agent traces leaked for fewer than 0.05% of users not in Privacy Mode; Privacy Mode prompts and code, production systems and customer repositories were untouched. The last attacker activity was blocked and Cursor considers the incident contained.
Affected: Anthropic.claude-code 2.1.312 or 2.1.313 · We had it 14 Nov, 23:04 UTC
Holes in AI tools
crissyfield.de · yesterday
Lovable
Takeaway: over 200,000 chat messages
Researchers at Crissy Field examined 159,041 Lovable projects and found sensitive data sitting wide open: 24,664 Supabase-backed apps served database records as soon as a page loaded, and with plain read-only requests over the same public API key the team pulled email-bearing records from 10,560 tables across 7,403 sites — no login and no break-in, just missing Row Level Security rules in the database.
Lovable's database integrations follow recurring patterns, so a method that works on one app works on thousands of others with little extra effort, and non-technical builders rarely spot the gap. If anyone in your organisation shipped a Lovable- or other vibe-coded app on Supabase, assume its tables need review before real customer data lands in them. One German lifestyle forum exposed more than 80,000 profiles and over 200,000 chat messages; a service provider exposed over 1,000 records carrying passport numbers. Lovable told Cybernews it sees a known industry-wide Supabase configuration problem rather than a flaw in its AI, and offers a security scan before launch.
bleepingcomputer.com · yesterday
FaceStealer
Takeaway: over $400 million in crypto
Not a product hole, but a documented case of attackers weaponising AI against the people who build software: the FBI says a North Korean campaign stole over $400 million in crypto by hijacking accounts on Slack, LinkedIn and Teams, luring victims into fake video meetings that played recordings of people the target knew, then pushing a bogus "technical fix" link that installed credential-stealing malware.
Open-source maintainers are explicit targets: the FBI describes a thwarted attempt against the maintainer of a JavaScript library with tens of millions of weekly downloads and says it confirmed further attempts against developers in the npm and Rust crate ecosystems — a supply-chain route into software most readers run. Verify unexpected meeting invites out-of-band, and never run a script offered as a fix for a meeting glitch. Arctic Wolf attributes the campaign to BlueNoroff, with 677 targets across 58 countries; U.S. authorities have seized the cloud infrastructure behind it.
Found or caused by AI
NLTK ConllCorpusReader DoS — A high-severity (CVSS 7.5) GitHub advisory covers a quadratic-time CPU denial of service in NLTK: any service that feeds untrusted CoNLL text to ConllCorpusReader.srl_instances() can be pinned by a crafted corpus — roughly 1 MB costs tens of seconds of single-core CPU, and a few megabytes reach minutes. Versions up to and including 3.10.3 are affected; 3.10.4 fixes it, and deployments that never call srl_instances() are unaffected. The advisory credits the report to "manus-use", filed by GitHub user zx (@manus-pi) — a handle and credit that, alongside a companion batch of algorithmic-DoS fixes against the same library, point to automated bug-hunting, though the advisory does not confirm it. Either way, agent-filed reports are landing on ordinary Python packages, and they deserve the same patch discipline as human ones. The advisory includes proof-of-concept timings: a 320-predicate block cost 10 seconds of CPU against 4.45 milliseconds for a single-predicate control.
Affected: <= 3.10.3 · Fixed in: 3.10.4 · Disclosed 14 Nov · We had it 14 Nov, 23:05 UTC
One thing to do
Turn on Row Level Security before an AI-built app holds real data. In the Supabase SQL editor, run `select tablename, rowsecurity from pg_tables where schemaname='public';` — any table with user data showing `rowsecurity = false` is readable by anyone holding the site's public API key. Enable it with `alter table <name> enable row level security;`, then add policies, because enabling RLS blocks all rows until a policy grants access. Run this against every app a team built with Lovable, Bolt, v0 or similar before it touches customer data. crissyfield.de
