Two security stories led the week. Cursor says a key leaked by one of its vendors ended up exposing secrets that customers sent to its support team. Separately, researchers found Lovable apps showing private data to logged-out visitors. Several price changes also landed, from Vercel builds to Gemini models and Codex speed tiers.

Coding Agents: Cursor’s investigation reaches support data

Cursor traced the fake Claude Code extension in its marketplace to a leaked Fireworks key. Fireworks provides model inference for Cursor. Cursor says an attacker used the exposed key to read stored coding agent traces, which are records of agent sessions. The attacker then used credentials found there to reach SaaS tools Cursor uses.

Cursor says some credentials that customers had sent to its support team were also accessed. If you ever pasted an API key or password into a support message, replace it. The same applies to a key pasted into an agent chat in 2025 with Privacy Mode off, because Cursor says its redaction of these traces was imperfect.

GitHub Copilot code review now flags exposed secrets in pull requests (proposed code changes awaiting review), in public preview. GitHub says it matches more than 500 known credential formats and uses AI models for the rest. A new Adaptive setting also picks Lite, Balanced, or Max review depth for each pull request. Removing a key from the final diff does not erase it from the branch history, so replace any key that reached a pull request.

Claude Code 2.1.323 added an experimental Code Mode, off by default. Claude writes one JavaScript program that makes many tool calls at once. Your deny rules and pre-action hooks still apply to each call. In auto mode, where a separate classifier decides what runs without asking you, the classifier reviews the script as a whole, not each call. Anthropic says it cannot see values computed while the script runs, so keep hard rules for calls you must control.

Codex Ultrafast writes code and replies up to 4.5x faster, but model speed is only part of a task. In OpenAI’s Android timer test, the average task fell from 16.7 seconds on Standard to 15.7 on Ultrafast, about 6%, because tests took roughly 13 seconds in every mode. Through the API, Ultrafast costs 3x Standard until February 28, 2027, then 6x from March 1. OpenAI has not said how it counts against Pro plan limits.

VS Code will remove its Local harness for Copilot agents in February 2027. The harness is the component that coordinates the model, context, and tools in a session. Start a new session with Copilot as the Session Target, then check that your saved customizations still work.

AI Builders: Exposed data, parallel copies, and budget caps

Crissy Field researchers examined 159,041 Lovable projects and found apps serving passport numbers, chat messages, and emails without a login. One forum exposed over 200,000 chat messages. The researchers found email addresses in 7,403 of 24,664 sites that loaded database records on page load. That is about 30% of that subset, not of all Lovable apps, and some addresses may be public contact details.

To check your own app, open it in a private browser window without logging in and see what loads. Then review who can read each database table.

Lovable Drafts copy an entire app, database included, so a teammate or client can try a change while the live app stays untouched. Each draft has its own chat and preview. Changes go live only after you accept the draft and publish. Lovable has not said which plans include drafts or whether they use extra credits.

Replit Missions coordinates a swarm of agents toward one large goal, such as replacing a legacy system. You set a spending cap and decide which outward-facing steps need approval. Work stops when the budget runs out, finished or not. Replit reports that its own test, building a replacement for an application monitoring product called DataCat, finished within a $3,000 budget and two days. Results on your project may differ, and Replit has published no pricing for Missions.

Infra & Deployment: Builds, model prices, and trial abuse

New Vercel Pro and Enterprise teams now pay for Standard builds by usage instead of getting one included slot. A build is the step that turns your code into a running site each time you deploy. On Pro the rate is $0.014 per build minute, so a five-minute build costs about seven cents before your monthly credit. Vercel’s changelog describes no change for existing teams.

Gemini 4 Neon costs $3 per million input tokens and $15 per million output tokens, where tokens are the chunks of text a model reads and writes. Carbon is half price at launch, at $2 and $10, so Neon is the most expensive of the three models today. According to Simon Willison, Carbon rises to $4 and $20 on December 1, which would make Neon 25% cheaper. If an app runs on Carbon, compare Neon on a few of your own tasks before the date.

Firebase Crashlytics for web entered public preview. It alerts you to new and returning JavaScript errors in users’ browsers, with native support for Next.js, React, and Angular. It requires the pay-as-you-go Blaze plan, and a busy app should check Firebase’s pricing first.

Claire Vo says abuse of free trials cost ChatPRD tens of thousands of dollars. Fraudsters farmed trials and promo codes for compute, and she ended trials. Any product that gives away model usage faces the same math, and a spending cap on the free tier limits the damage.

GitHub also reported that some pull request merges timed out on November 16, with a peak of 7.1% of attempts at 15:17 UTC. GitHub says overloaded storage servers caused it and that the incident is resolved. Git pushes and fetches stayed available, with higher latency.


End of article