Releases Index

BetaWeekly digests are a beta — we're trying something new. Feedback welcome.

Cloudflare opens its agent browser and tightens the cost and security levers

November 9–15, 2026

CloudflareVercelRailway
Cloudflare, Vercel, Railway

Cloudflare open-sourced Kitesurf, its Workers-native browser engine for AI agents, while adding account spending limits for Workers and Durable Objects and wiring GitHub secret scanning into Access service token revocation. Cloudflare's runtime and SDKs also got a quieter but broad set of changes: local-first Artifacts bindings, new Workflows support, and a hardening pass across AI SDK request validation.

The week at a glance

Replay the week
FeaturesEnhancementsFixesMajorLandmarkTile size reflects impact

Cloudflare's big swings: an open agent browser and spend controls

The headline of the week is Kitesurf going open source under Apache 2.0. The engine parses HTML, applies CSS, runs page JavaScript, and renders screenshots entirely on Workers — no browser container — and speaks the Chrome DevTools Protocol, so existing automation tooling can drive it. It's still experimental and ships without built-in authentication, and deploying it needs Workers Paid for Dynamic Workers, but as a self-hostable browser for agents it's a substantial piece of infrastructure to hand to the community.

Cost control got its own story. Workers and Durable Objects now support spending limits, set per product in the dashboard or via the cf CLI, that block supported operations once a threshold is hit, with optional email alerts that leave applications running. It's a deliberate pairing with AI Gateway's per-user and per-model budgets, giving teams one place to cap compute and inference spend together.

Security that closes the leak paths

Cloudflare Access now disables service tokens that GitHub detects in public repositories; secret scanning reports exposed Client Secrets and Cloudflare kills the token and emails admins, with no rotation grace period — the old secret is discarded immediately, so plan for a fast rotate. Other scanners can match the cfast_-prefixed format with a custom rule.

Network edge hardening continued: Tunnel's URL normalization extends to semicolon-delimited path parameters, closing a class of bypass for ingress rules that rely on Access validation, though the path sent to the origin is unchanged. It stays off by default and becomes default for new cloudflared releases starting 2027-02-08. The macOS One Client bumped its floor to macOS 15 while adding automatic Global Acceleration endpoint selection, an in-client Diagnostics page, and fixes for wake-from-sleep DNS failures and a managed registration dropped when the MDM config file briefly vanished.

The Workers toolchain goes local-first

Across Wrangler, the Vite plugin, Miniflare, and Workers Utils, Artifacts bindings now default to the local simulator, matching KV, R2, and D1 — set remote: true to point at deployed repos (local Artifacts needs Git 2.32+). The same config work rejects limits, concurrency, and schedules on Workflow bindings declared via exports, and streaming tail consumers moved to experimental.streamingTailConsumers, with a codemod to migrate. Wrangler also fixed relative source maps in experimental Build Output, and the vitest plugin now prepares and runs Containers from cloudflare.config.ts, which means Docker for those tests. Underneath, workerd added Queues in standalone and cluster mode plus a batch of Node compatibility, and then brought Workflows to standalone workerd while fixing cron double-runs on skewed clocks.

The platform surface widened too: AI Search can index D1 databases as a source, D1 exports databases containing virtual tables with a CLI-only compatibility mode, and RealtimeKit auth tokens gained custom validity windows and single-token revocation instead of a fixed 100-day life. On the model side, Workers AI added the 594B-parameter multimodal Inkling 2 Small and EmbeddingGemma 2 for cross-modal search.

AI SDK hardens validation while Vercel widens model access

The AI SDK's week was mostly about not falling over on hostile input: malformed UI messages no longer exhaust memory because each part is validated only against its own type's schemas and validation stops at the first failure, and streaming long arrays with Output.array no longer burns CPU since completed elements are parsed incrementally. On the transport side, MCP moved connection-time DNS validation into the optional guarded-fetch package, and remote file URLs now require an explicit download hook rather than silently fetching through global fetch. Elsewhere, OpenAI's provider preserves prompt cache across structured output schema changes, the WorkflowAgent gained toolApproval and toolOrder, and ScaleDown joined as a provider. Vercel's AI Gateway picked up both Inkling 2 Small and Gemini 4 Hydrogen, and Turborepo made turbo.lock creation roughly 10x faster while scoping saved Vercel credentials to Vercel endpoints. Railway's CLI, meanwhile, added background agent tasks and a ca task command.

AI-generated digests may contain mistakes.
Releases covered28
Cloudflare
Vercel