Cursor now lets a cloud agent take an internal app from a description to a published link, so a non-engineer can ship a tool without handing it off. Vercel spent the week on checks around what gets deployed: it blocks flagged npm packages, reviews every push to a connected GitHub repository, and lets throwaway projects delete themselves.

AI Builders: Cursor agents can publish an app behind team sign-in

Cursor’s changelog says you start a cloud agent, an agent that runs on Cursor’s servers, and tell it you want an app. It builds the web app, adds a Postgres database if needed, sets up login through Cursor sign-in, and arranges hosting. At publish time you pick a domain and get a link. Our full write-up covers the details.

Hosting, login and a database used to be separate jobs, which is often why an agent-built tool never left its author’s laptop. Now the agent handles them together. Before publishing, Cursor runs security reviews for problems such as secrets written into app files, and a blocked publish needs approval. Team admins can see every app and secret, add their own review rules, and unpublish any app.

The person describing the app also decides what it connects to, because the agent asks for the secrets it needs. Admins should look at which secrets people hand to these apps. Builders should grant only the access an app needs.

Infra & Deployment: Vercel checks dependencies and code, and cleans up after demos

Three Vercel changes cover different points in the same pipeline.

Blocked packages now fail builds. Vercel checks dependencies against a blocklist before and during install on all plans. A flagged version, whether malicious or merely vulnerable, stops the deployment, and the error names the package to update. A project you have not touched can fail on its next build if one of its versions was added to the list since.

Security Review checks every push. Vercel Security Review is in public beta for Pro and Enterprise teams. Once enabled, it reviews every push to a connected GitHub repository, including direct pushes to production, and lists findings in the Vulnerabilities dashboard.

At launch, Vercel said reviews would be free until November 10. It announced pricing at the model provider’s token rates plus $0.25 per million tokens. Each push triggers a review, so cost follows how often you push. The scanner analyzes source code only and does no runtime testing.

Throwaway projects can expire. Vercel CLI, the command-line tool for deploying from your terminal, now has vercel deploy --delete-project-after 1d in version 65.12.0 or later. It deletes the whole project, not just a deployment, after a time you choose from one hour to 30 days. The flag works only when the deploy creates the project, and vercel project keep cancels the timer.

Vercel also made tracing generally available on all plans. A trace is a record of the steps a request took, with the time each one cost. You can search traces in the dashboard or CLI, and pricing is unchanged from the beta. Starting with Next.js 16.5, new projects that include a Node.js function get tracing after their first successful build, using a rule that traces 100% of requests. To opt out, remove all sampling rules in Settings, then Tracing.

A GitHub incident affected Actions and Pages on November 4. Actions runs and status updates were delayed by up to 25 minutes in two windows, and Pages deployments fell to about 4% of normal volume. GitHub traced it to a configuration change that paused several event streams, and says it has resolved the incident. Vercel’s dashboard and builds were delayed on November 5, also now resolved. If a deploy hangs during an incident like these, check the status pages before changing your code.

Coding Agents: Cheaper subagents, and Copilot in more editors

Claude Code 2.1.316 has Haiku 5.5, Anthropic’s small, low-cost model, write the status line for background subagents, the helpers that keep working while you do something else. Anthropic says sessions with background subagents use about 5% less of the usage limit. Sessions without them see no change.

GitHub Copilot reached more places. It now works in JetBrains Air, where you delegate coding tasks to agents, using your existing Copilot subscription. Cloud tasks there need a JetBrains Air Teams business organization. It is also in public preview in Xcode’s coding assistant starting with the Xcode 27.3 beta on macOS Tahoe 26.6 or later. @GitHub in Slack and Teams is now generally available, so a chat thread can become an issue or pull request.

OpenAI introduced GPT-6 Cyber, a model for security research. OpenAI says it scored 48.5% on ExploitGym, a test of whether a model can exploit a specified vulnerability, up from 32.1% for GPT-5.6 Cyber. OpenAI also says it helped find 81 previously unknown high-severity Firefox vulnerabilities. Interested teams apply through OpenAI’s Daybreak program.

Workflow: Agents build vague tickets literally

Product consultant Rich Mironov argues that once agents can clear a backlog, vague customer requests become the risk. His examples: a request for a new sort option could force expensive database work, and a shortcut in account creation could break compliance rules. In his words, “Tickets are wishes, not specs, and agents grant wishes literally.”

His fix is to have an agent screen each request against written product, regulatory and performance constraints, then spend half an hour a week reviewing what it rejected. The screen only catches constraints someone has written down, and the review covers rejections, not approvals that slipped through. A team adopting this should also spot-check a sample of approved requests until the rules have a track record.


End of article