Skip to content

Changelog

New updates and improvements at Cloudflare.

Run custom scripts in Browser Run with /script

Browser Run now runs custom browser automation written with Playwright, Puppeteer, or the raw Chrome DevTools Protocol (CDP) in a single API request, with the new /script endpoints. Submit a JavaScript function, and Browser Run executes it in an isolated Worker connected to a browser session, so you do not need to deploy a Worker.

curl -i -X POST "https://api.cloudflare.com/client/v4/accounts/<accountId>/browser-run/script" \
  -H "Authorization: Bearer <apiToken>" \
  -H "Content-Type: application/json" \
  -d '{
    "code": "async ({ page }) => { await page.goto(\"https://example.com\"); return { title: await page.title() }; }",
    "library": "playwright"
  }'
HTTP/2 200
content-type: application/json
cf-browser-session-id: 8f7c2e3a-1b4d-4c6e-9a0f-2d5b7e9c1a3f
cf-browser-target-id: A1B2C3D4E5F60718293A4B5C6D7E8F90

{
  "success": true,
  "result": {
    "title": "Example Domain"
  }
}
  • POST /script acquires a browser, runs your function, and returns its result as { "success": true, "result": ... }. The session and page are returned in the cf-browser-session-id and cf-browser-target-id response headers (except Kitesurf runs).
  • POST /script/<sessionId> runs follow-up scripts in a session you already hold, from /script or from POST /devtools/browser. Pass targetId to attach to a specific page, or create extra pages to run scripts concurrently in one session.
  • Session launch options (keep_alive, recording, lab) go in the /script query string. Use browser=kitesurf to run a CDP script on Kitesurf.
  • From a Worker, call the same endpoints through the browser binding with env.BROWSER.script(), without an API token.
  • From the Cloudflare dashboard, write and run scripts against a live browser in Script Studio, without an API token.
  • Neither endpoint closes the session after your script runs. Close it with DELETE /devtools/browser/<sessionId> or browser.close(). Otherwise, it ends at its idle timeout.

For more information, refer to Custom scripts.

View delivery analytics for hosted images

Cloudflare Images lets you store, optimize, and deliver hosted images through a scalable image pipeline. The new Images Analytics dashboard ↗︎ provides up to 30 days of delivery data for assets stored in Images, helping you understand how and where they are requested.

Hosted images analytics showing request totals, countries, and response status codes

In Images & Stream > Hosted images > Analytics, you can:

  • Investigate traffic trends and sources. Review request volume over time, then filter by country, autonomous system number (ASN), or referrer to trace spikes and unexpected traffic.
  • Troubleshoot delivery. Break down requests by response status code to identify errors. Filter by browser or operating system to understand which output formats different clients receive.
  • Understand asset and variant usage. Find the most-requested source images and compare predefined variants with inline configurations to see which configurations account for the most requests.

Use cf sql query to query delivery data from the terminal. The SQL API exposes this data in the events.imagesHostedRequests dataset:

cf sql query \
  'SELECT httpStatus, COUNT(*) AS requests
   FROM events.imagesHostedRequests
   GROUP BY httpStatus
   ORDER BY requests DESC
   LIMIT 20' \
  --scope-account "<ACCOUNT_TAG>" \
  --time-since "<START_TIME>"

The SQL API also exposes transformation data, such as binding requests, from the events.imagesTransformations dataset:

cf sql query \
  'SELECT isBindingRequest, scriptName, originHost, COUNT(*) AS transformations
   FROM events.imagesTransformations
   GROUP BY isBindingRequest, scriptName, originHost
   ORDER BY transformations DESC
   LIMIT 5' \
  --scope-account "<ACCOUNT_TAG>" \
  --time-since "<START_TIME>"

Open the Images Analytics dashboard ↗︎ or use the Cloudflare CLI to investigate recent traffic changes and identify the hosted images and variants driving requests.

CVE pages in Zero Wire

Cloudflare Radar Zero Wire ↗︎ now has a page for every published CVE. Each page brings together what you need to decide how urgently to act on a vulnerability: how severe it is, whether it is being exploited, what it affects, and how to fix it.

Animation searching the CVE list for log4j, filtering to Critical severity, sorting by highest score, and opening the Log4Shell CVE page, then scrolling through all of its sections, from exploitation to references

Browse CVEs

The CVE list ↗︎ includes every published CVE. Sort them by date or CVSS score, search by CVE ID or keyword, such as a product or vendor name, and filter by severity and year. Each row shows the CVE's score and highlights CVEs in the CISA Known Exploited Vulnerabilities (KEV) catalog and those with a fix available.

CVE pages

Each CVE page, such as CVE-2021-44228 (Log4Shell) ↗︎, starts with its CVSS score, description, and key dates, followed by:

  • Exploitation: whether the CVE is in the CISA KEV catalog, including known ransomware use, the CISA SSVC assessment, and the FIRST EPSS probability of exploitation.
  • Affected products: the affected vendors, products, and versions, as listed by the organization that published the CVE and by NVD.
  • Recommended actions: how to fix the vulnerability, or reduce the risk until you can, including CISA's guidance for known exploited CVEs.
  • Severity: every CVSS score for the CVE, with each metric explained.
  • Weaknesses and attack patterns: the related CWE weaknesses and CAPEC attack patterns.
  • References: advisories, patches, and other sources.

CVEs cited in Zero Wire briefings ↗︎ link to their CVE pages, and each CVE page links back to the briefings that cite it.

Search and follow

Search for a CVE ID or keyword from anywhere on Radar, using the search bar or Cmd+K (Ctrl+K on Windows and Linux), to jump straight to a CVE page.

Radar search showing CVE results for log4j

To follow new CVEs, subscribe to the RSS ↗︎, Atom ↗︎, or JSON ↗︎ feed. Each feed can be filtered by severity, such as the feed of Critical CVEs ↗︎.

Explore the CVE list ↗︎ on Cloudflare Radar.

Dispatch namespaces can no longer be renamed

You can no longer rename a dispatch namespace after creating it.

Renaming a namespace could cause API requests to read, deploy, or delete Workers to fail because dispatch Workers and outbound Workers retained the namespace name they were deployed with. This change prevents those failures by rejecting rename requests.

What changes:

  • PUT and PATCH requests to the dispatch namespace API return a 400 error with code 100413 if they change name:

    Dispatch namespaces cannot be renamed. The name in the request must match the namespace's current name. To use a different name, create a new dispatch namespace.
  • wrangler dispatch-namespace rename returns the same 100413 error.

  • Terraform applies that attempt to change name on a cloudflare_workers_for_platforms_dispatch_namespace resource fail.

To move your Workers to a namespace with a different name:

  1. Create a new dispatch namespace with the name you want.
  2. Deploy your user Workers to the new namespace.
  3. Update the dispatch namespace binding in your dispatch Worker to reference the new namespace, then redeploy the dispatch Worker.
  4. Verify that requests are routed to the new namespace and your Workers work as expected.
  5. Delete the user Workers in the old namespace, then delete the old namespace.

For more information, refer to How Workers for Platforms works.

Capture PDFs, links, scraped elements, and extracted JSON with /snapshot

Browser Run's /snapshot endpoint can now return four more formats from a single page load: pdf, links, scrape, and json. Before, /snapshot returned HTML, screenshots, Markdown, and the accessibility tree, so a workflow that also needed a PDF or structured data had to prepare the same page again in a second request.

Use the new formats when you need to:

  • PDF: print the page, with the same pdfOptions as the /pdf endpoint.
  • Links: collect the links on the page, with the same visibleLinksOnly and excludeExternalLinks options as /links.
  • Scraped elements: return the elements matched by elements, the same payload as /scrape. elements is required when you request it.
  • Extracted JSON: extract structured data from the page HTML with Workers AI, with the same prompt, response_format, and custom_ai fields as /json. prompt or response_format is required, and the request incurs Workers AI usage.

Request them through formats, the same way as the existing formats. This example returns a PDF, the page links, and the text of each paragraph:

const response = await env.BROWSER.quickAction("snapshot", {
	url: "https://example.com/",
	formats: ["pdf", "links", "scrape"],
	elements: [{ selector: "p" }],
});

const { result } = await response.json();
console.log(result.pdf);
console.log(result.links);
const response = await env.BROWSER.quickAction("snapshot", {
	url: "https://example.com/",
	formats: ["pdf", "links", "scrape"],
	elements: [{ selector: "p" }],
});

const { result } = await response.json();
console.log(result.pdf);
console.log(result.links);

Formats are captured in a fixed order rather than the order of the formats array, and the PDF is printed last because print handlers can change the page. A live page can change between captures, and any format that fails fails the whole request. For JSON snapshots, one actionTimeout budget (default 60 seconds, maximum 120 seconds) covers navigation, captures, retries, and the AI request.

You still need to request at least two distinct formats. If you only need one, use the corresponding single-format endpoint.

Refer to the /snapshot documentation for the full list of accepted values and options.

Improved Durable Object Reliability During Runtime Updates

Cloudflare has improved how Durable Object storage handles machines restarting during runtime updates.

Each write to a Durable Object's storage must be confirmed by a majority within a group of machines for durability. Machines now proactively remove themselves from the group when restarting, so that they can be replaced before the restart happens.

Previously, this could take up to a few minutes. During that time, a write could fail to reach a majority, and the object was reset with Internal error in Durable Object storage caused object to be reset, disconnecting its WebSockets.

This can make these resets less frequent for SQLite-backed Durable Objects, especially ones that write to storage only occasionally, such as those using WebSocket hibernation. Actual results depend on your workload and traffic patterns.

The change is live in all regions. No action is needed by customers to see benefits from this change. Some disconnects during runtime updates are still expected, so clients should continue to reconnect when a WebSocket closes.

To learn more, refer to Use WebSockets.

Apply Gaussian sharpening with Images

Cloudflare Images lets developers resize, adjust, and manipulate images at scale through a broad set of optimizations, including image adjustments such as sharpen, brightness, contrast, and gamma.

The sharpen parameter now supports Gaussian sharpening, in addition to the existing box-filter sharpening. Sharpening makes fine details appear crisper by increasing the pixel-to-pixel contrast along edges, where adjacent pixels differ in brightness or color. Scalar values such as sharpen=5 continue to use the box filter.

Different images benefit from different types and levels of sharpening. Text may need strong edge definition for readability; product photography often uses a subtler effect to avoid emphasizing grain; and illustrations with bold, clean lines and large areas of solid color can benefit from emphasizing larger edges and object outlines.

Both box and Gaussian sharpening use unsharp masking, a technique that first creates a blurred version of the source image and subtracts it from the original to isolate edge details. Sharpening then adds a controlled amount of that detail back to the source along these boundaries, improving the perceived clarity of an image.

Each filter uses a different method to produce the blur used for this comparison. The existing box filter produces a uniform blur by averaging each pixel with its eight surrounding pixels, weighing all nine equally. A Gaussian filter gives more weight to pixels closer to the center and provides more control to help prevent sharpening from treating unintended parts of an image as edges.

Box sharpening offers a simple, predictable configuration for direct, crisp effects on text, screenshots, and simple graphics. Gaussian sharpening can produce smoother transitions for photography and offers more control over which details are emphasized.

To use Gaussian sharpening, set the mode setting to gaussian, then configure its behavior using the additional amount, sigma, and threshold settings:

  • Amount adjusts how strongly Images increases contrast around identified edges. It accepts values from 0 to 10.
  • Sigma controls the spread of the Gaussian blur, which determines how far the sharpening effect extends around an edge. It accepts values greater than 0 and up to 5.
  • Threshold (optional) excludes pixel differences below a specified value, which can reduce sharpening of low-contrast noise, grain, or smooth areas. It accepts integers from 0 to 255.

Together, these settings let you tailor the sharpening effect to your content.

In the URL interface, these settings are grouped beneath the sharpen parameter:

sharpen=(mode=gaussian,amount=1,sigma=1.5,threshold=0)

The same structure maps directly to the cf.image object in a Workers fetch() subrequest:

cf: {
	image: {
		sharpen: {
			mode: "gaussian",
			amount: 1,
			sigma: 1.5,
			threshold: 0,
		},
	},
}

When using the Images binding, pass the structured sharpen object directly to .transform():

const result = await env.IMAGES.input(source.body)
	.transform({
		sharpen: {
			mode: "gaussian",
			amount: 2,
			sigma: 1.5,
			threshold: 0,
		},
	})
	.output({ format: "image/avif" });

return result.response();

Refer to sharpen.

View R2 operations as requests per second

You can now view Class A and Class B operations as requests per second (RPS) in addition to request count on the Metrics page for an R2 bucket in the Cloudflare dashboard.

Go to Bucket metrics ↗

Use the Request count and Requests/second toggle on the Class A Operations and Class B Operations charts to switch between views.

Request count

Request count shows the total number of operations in each time interval. Use it to understand overall usage and how operations contribute to your bill.

R2 Class A operations chart showing request count per hour

Requests per second

Requests/second shows the rate of operations over time. Use it to identify traffic peaks.

R2 Class A operations chart showing average requests per second

For more information, refer to R2 metrics and analytics.

Autonomous system types on Cloudflare Radar

Cloudflare Radar now shows the type of each autonomous system (AS): Access ISP, Hosting/Cloud/Content, or Organization. AS types are inferred automatically, so an individual AS can be misclassified. They make it easier to tell, for example, whether traffic or attacks come from residential and mobile networks or from data centers.

  • Access ISP (ACCESS_ISP): networks that connect end users, such as home broadband and mobile operators, plus transit providers.
  • Hosting/Cloud/Content (HOSTING_CLOUD_CONTENT): hosting providers, cloud platforms, and content delivery networks.
  • Organization (ORGANIZATION): enterprises, universities, governments, banks, and IXP route servers.

ASes without a type are shown as Unclassified. To report a wrong AS type, email radar@cloudflare.com.

Chart of HTTP traffic from the top five ASes beside a table of the top ten, with house, cloud, and building icons before the AS numbers

Top ASes by HTTP traffic in Portugal ↗︎, with the type of each AS shown before its number.

Radar pages

  • AS pages, such as AS7922 ↗︎, show the AS type in the AS information card.
  • An icon before the AS number shows its type across Radar. Hover over the icon to see the AS type.
  • The Data Explorer ↗︎ can break down and filter HTTP, NetFlows, DNS, and layer 7 attack data by AS type.
AS information card listing the AS type as Access ISP, above a table of related ASes with house, cloud, and building icons

The AS type in the AS7922 ↗︎ information card.

Line chart of HTTP requests by AS type, scaled from zero to the maximum, where Access ISP stays well above Hosting/Cloud/Content and Organization stays near zero. A bar below shows Access ISP at 66% and Hosting/Cloud/Content at 33% of requests

Worldwide HTTP requests by AS type over the last four weeks, from the Data Explorer ↗︎.

API

  • The Get AS details, List ASes, and Get AS details by IP address endpoints return a new category field with the AS type, or null when the AS is not classified.
  • The HTTP, NetFlows, DNS, and layer 7 attacks endpoints accept a new asCategory filter, and their summary and timeseries groups endpoints support a new AS_CATEGORY dimension. Both accept ACCESS_ISP, HOSTING_CLOUD_CONTENT, ORGANIZATION, and UNCLASSIFIED, where UNCLASSIFIED also covers traffic that could not be mapped to an AS. For HTTP, they are supported for date ranges starting on or after 2023-05-01.
  • The global search endpoint returns an entity with details of the matched AS, including its category, for AS results, and of the matched region for region (adm1 and adm2) results.

For example, to get the share of HTTP requests by AS type in the last week:

Get HTTP requests summary by dimensionbash
curl "https://api.cloudflare.com/client/v4/radar/http/summary/AS_CATEGORY?dateRange=7d" \
	--request GET \
	--header "Authorization: Bearer $CLOUDFLARE_API_TOKEN"

Visit an AS page ↗︎ on Cloudflare Radar to see its type, or read the AS type definitions.

Redesigned Spectrum dashboard

The Spectrum pages in the Cloudflare dashboard have been redesigned. Your existing applications and their settings do not change.

The Spectrum Applications page lists each application's type, edge port, origin, and edge IP type, with the last minute of connections, ingress, and egress. You can search the list and sort it by column.

Creating or editing an application now opens a full page instead of a modal, with settings grouped into Application Configuration, Origin Configuration, and Advanced Options. To edit an application, select its hostname, or select the three dots next to it > Edit.

You can now create and edit applications that use static IPs or BYOIP addresses in the dashboard. Under Edge IP Type, select Dedicated IPs. Leave Subdomain blank to create an application on the zone apex.

New TCP and UDP applications created in the dashboard now have IP Access Rules turned off by default, which matches the API default. To enforce IP Access rules for an application, turn on IP Access Rules under Advanced Options. Existing applications keep their current setting.

To create your first application, refer to Get started with Spectrum.

To share feedback on the redesigned dashboard, complete this form ↗︎.

B-frame detection for Low-Latency HLS live streams

Cloudflare Stream now emits a warning when B-frames are detected in Low Latency HLS live broadcasts.

The most common cause of freezing and stuttering in LL-HLS live streams is when encoders include B-frames. These frames can reference content that has not yet been transmitted, stalling preparation for delivery. Therefore, B-Frames are not supported when LL-HLS is enabled for a live input.

Stream sends the warning ERR_LIVE_INPUT_LLHLS_INCOMPATIBLE once per live session if applicable. It is reported alongside other error and status events in:

To resolve the warning, turn off B frames in your broadcast application settings. Alternatively, turn off Low-Latency HLS in your live input settings, but playback latency will be higher.

For more information, refer to Stream live troubleshooting.

WAF Emergency Release - 2026-11-19

This release introduces updates to an existing PHP Remote Code Execution rule to improve detection accuracy, as well as a few new generic, PHP and Ruby CVE specific rules. The new PHP rules are being deployed in log mode by default (in lieu of False Positive risk) and can be overridden after evaluating requests matching the rules. Supplementary information around the CVEs are available through Zero Wire phpAfterShell and six other PHP critical flaws ↗︎ and ruby2shell: heap overflow in Ruby string case conversion ↗︎.

RulesetRule IDLegacy Rule IDDescriptionPrevious ActionNew ActionComments
Cloudflare Free RulesetN/APHP Remote Code Execution Attack (CVE-2026-109618 ↗︎, CVE-2026-109619 ↗︎, CVE-2026-115560 ↗︎, CVE-2026-126714 ↗︎) #6N/ABlockRule detection and description updated.
Cloudflare Free RulesetN/ARuby - Heap Overflow - CVE:CVE-2026-13282 ↗︎N/ABlockThis is a new detection.
Cloudflare Free RulesetN/ARuby - Integer Overflow - GHSA-87fm-7w84-j2pxN/ABlockThis is a new detection.
Cloudflare Managed RulesetN/APHP Remote Code Execution Attack (CVE-2026-109618 ↗︎, CVE-2026-109619 ↗︎, CVE-2026-115560 ↗︎, CVE-2026-126714 ↗︎) #6N/ABlockRule detection and description updated.
Cloudflare Managed RulesetN/ARuby - Heap Overflow - CVE:CVE-2026-13282 ↗︎N/ABlockThis is a new detection.
Cloudflare Managed RulesetN/ARuby - Integer Overflow - GHSA-87fm-7w84-j2pxN/ABlockThis is a new detection.
Cloudflare Managed RulesetN/APHP-FPM - CVE:CVE-2026-129289 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/APHP SOAP - CVE:CVE-2026-15353 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/APHP SOAP - CVE:CVE-2026-129506 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/AGeneric Rules - Authentication BypassN/ABlockThis is a new detection.
Cloudflare Managed RulesetN/AAnomaly - Embedded Application RequestN/ABlockThis is a new detection.

Inbound TCP and gRPC for Workers and Containers now in public beta

Inbound TCP connections to Workers and bidirectional gRPC in Containers are now in public beta.

Worker-origin Spectrum applications are available on Pro and Business zones, and on Enterprise zones with Spectrum Enterprise enabled. Without Spectrum Enterprise, Pro and Business zones support one Worker-origin application per zone, on one TCP port.

flowchart LR
    Client[TCP client] <--> Spectrum
    Spectrum <--> Worker["Worker connect()<br/>Handle or forward bytes"]
    Worker <--> DO["Durable Object<br/>connect()"]
    DO <--> Container["Container<br/>gRPC or other TCP server"]

Forward a Worker socket to a Durable Object

Spectrum sends public TCP connections to your Worker's connect() handler. The Worker opens a socket to the GrpcContainer Durable Object and pipes bytes in both directions:

src/index.jsjs
export default {
	async connect(socket, env) {
		const container = env.GRPC_CONTAINER.getByName("grpc-demo");
		const durableObjectSocket = container.connect("grpc-container:50051", {
			allowHalfOpen: true,
			secureTransport: "off",
		});

		try {
			await durableObjectSocket.opened;
			await Promise.all([
				socket.readable.pipeTo(durableObjectSocket.writable),
				durableObjectSocket.readable.pipeTo(socket.writable),
			]);
		} finally {
			await Promise.allSettled([socket.close(), durableObjectSocket.close()]);
		}
	},
};
src/index.tsts
export default {
	async connect(socket: Socket, env: Env): Promise<void> {
		const container = env.GRPC_CONTAINER.getByName("grpc-demo");
		const durableObjectSocket = container.connect("grpc-container:50051", {
			allowHalfOpen: true,
			secureTransport: "off",
		});

		try {
			await durableObjectSocket.opened;
			await Promise.all([
				socket.readable.pipeTo(durableObjectSocket.writable),
				durableObjectSocket.readable.pipeTo(socket.writable),
			]);
		} finally {
			await Promise.allSettled([socket.close(), durableObjectSocket.close()]);
		}
	},
} satisfies ExportedHandler<Env>;

Forward the Durable Object socket to its Container

The Durable Object receives the connection in its own connect() handler. Once its Container is running, getTcpPort(50051).connect() opens a socket to the gRPC server inside it:

src/index.jsjs
import { DurableObject } from "cloudflare:workers";

export class GrpcContainer extends DurableObject {
	async connect(socket) {
		const containerSocket = this.ctx.container
			.getTcpPort(50051)
			.connect("10.0.0.1:50051", {
				allowHalfOpen: true,
				secureTransport: "off",
			});

		try {
			await containerSocket.opened;
			await Promise.all([
				socket.readable.pipeTo(containerSocket.writable),
				containerSocket.readable.pipeTo(socket.writable),
			]);
		} finally {
			await Promise.allSettled([socket.close(), containerSocket.close()]);
		}
	}
}
src/index.tsts
import { DurableObject } from "cloudflare:workers";

export class GrpcContainer extends DurableObject<Env> {
	async connect(socket: Socket): Promise<void> {
		const containerSocket = this.ctx
			.container!.getTcpPort(50051)
			.connect("10.0.0.1:50051", {
				allowHalfOpen: true,
				secureTransport: "off",
			});

		try {
			await containerSocket.opened;
			await Promise.all([
				socket.readable.pipeTo(containerSocket.writable),
				containerSocket.readable.pipeTo(socket.writable),
			]);
		} finally {
			await Promise.allSettled([socket.close(), containerSocket.close()]);
		}
	}
}

allowHalfOpen: true preserves TCP half-closes. The Worker and Durable Object forward encrypted bytes; the Container terminates TLS and handles HTTP/2 and gRPC.

Run bidirectional gRPC in the Container

Run an existing gRPC server in any language, including services for real-time voice applications. This Node.js snippet shows the streaming handler and service registration. It sends a greeting, echoes incoming messages, and sends a final response when the client finishes:

container/server.jsjs
const path = require("node:path");
const grpc = require("@grpc/grpc-js");
const protoLoader = require("@grpc/proto-loader");

const definition = protoLoader.loadSync(
	path.resolve(__dirname, "../proto/bytes.proto"),
);
const proto = grpc.loadPackageDefinition(definition).cloudflare.grpcdemo;

function chat(call) {
	let count = 0;
	call.write({
		payload: Buffer.from("hello from the container gRPC server\n"),
	});
	call.on("data", ({ payload }) => {
		count += 1;
		call.write({
			payload: Buffer.concat([
				Buffer.from(`container echo ${count}: `),
				payload,
			]),
		});
	});
	call.on("end", () => {
		call.write({
			payload: Buffer.from("goodbye from the container gRPC server\n"),
		});
		call.end();
	});
	call.on("error", (error) => console.error("gRPC stream error", error));
}

const server = new grpc.Server();
server.addService(proto.ByteStream.service, { Chat: chat });

Follow the complete gRPC example for the protobuf definition, bindings, Spectrum routing, Secrets Store TLS credentials, and Container startup and readiness checks.

Workers and Pages now keep versions for 30 days

Cloudflare now keeps versions of your Workers and deployments of your Pages projects for 30 days. After 30 days, they are deleted, except your 100 most recent and what is currently deployed.

What is not affected:

  • Running applications: Deployed Worker versions and your current Pages production deployment are always kept
  • Rollbacks: You can still roll back to any of your 100 most recent Worker versions, and to any Pages production deployment that has not been deleted
  • Worker Previews: Worker Preview URLs always serve the latest deployment of that Preview. Use Worker Previews for branches and pull requests instead of Version URLs, which run with your production bindings

What you will notice:

  • Deleted versions and deployments no longer appear in the dashboard or API
  • The URL associated with each deleted Worker version, known as its Version URL (including any --preview-alias), stops working, as does the preview URL of each deleted Pages deployment

For more information, refer to Workers version retention, Pages deployment limits, and Compare workflows.

Automatically retrieve SAML IdP signing certificates

Cloudflare Access can retrieve signing certificates from a SAML identity provider's metadata endpoint. Your IdP can publish replacement certificates without requiring you to upload each one to Access.

In your generic SAML identity provider configuration, set Signing certificate source to Fetch certificates from IdP metadata and enter the IdP metadata URL. The endpoint must return SAML metadata XML over public HTTPS, without authentication or redirects. It must not return a JSON Web Key Set (JWKS).

For API configurations, use metadata_url instead of manual certificate fields:

{
	"type": "saml",
	"config": {
		"issuer_url": "https://idp.example.com",
		"sso_target_url": "https://idp.example.com/sso",
		"metadata_url": "https://idp.example.com/metadata.xml"
	}
}

Access caches the certificates and does not poll the endpoint.

Removing a compromised certificate from metadata does not immediately revoke the cached certificate or existing Access sessions. For setup, planned rotation, and emergency replacement guidance, refer to Discover IdP signing certificates.

Attach images, PDFs, and text files to Agent Lee

You can now attach files to a message in Agent Lee. Share a screenshot of an error, a PDF, or a log, HAR, or configuration file, and Agent Lee reads it as part of your question.

Previously, you had to describe what you were seeing or paste file contents into the chat by hand. A screenshot of a dashboard error or a long log was hard to get across that way.

How it works

  • Select the paperclip next to the message box, paste an image, or drag files onto the chat panel.
  • Each file appears as a chip above the message box. Remove any you did not mean to add, then send your message.
  • Agent Lee reads images directly, converts PDFs to text and reads them from the start, and reads text files in full or, for long files, from the start and the end.

Example requests:

  • "Here is a screenshot of the error I get when I open my site. What is wrong?"
  • "This HAR file is from a slow page load. Which requests are taking the longest?"
  • "Here is the error log from my Worker. What is causing these failures?"

Good to know

  • Supported files. PNG, JPEG, and WebP images up to 3.5 MiB, PDFs up to 5 MiB, and text files (.txt, .log, .json, .har, .csv, .yaml, .yml, .toml, .md, .xml) up to 100 MiB. You can attach up to four files per message, up to 14 MiB in total after trimming.
  • Text files are trimmed in your browser. A text file over 1 MiB (2 MiB for JSON, HAR, YAML, TOML, and XML) keeps its first and last lines. HAR files are always compacted before sending: request and response bodies and cookies are removed, and credential headers such as Authorization and Cookie are replaced with [redacted].
  • Scanned PDFs are not read. Agent Lee reads the text in a PDF. A PDF that is only scanned images, with no text layer, cannot be read.
  • Older attachments age out. Agent Lee keeps attachments from your most recent messages in view. Attach a file again to bring an older one back into the conversation.
  • Attachments are stored with the conversation. When an attachment ages out, it is removed from the stored conversation. Do not attach files that contain passwords, API tokens, or other secrets. For how Agent Lee uses your data, refer to Data usage.

Agent Lee remains in beta. Features and behaviors may change as the product evolves.

For details, refer to Agent Lee.

Hyperdrive-connected PlanetScale databases now available to Cloudflare Startup Program customers

Customers in the Cloudflare Startup Program ↗︎ can now create PlanetScale Postgres and MySQL databases from Cloudflare, with usage billed through their Cloudflare account.

Startup credits applied to PlanetScale usage is limited for each tier. PlanetScale usage beyond a tier's limit is billed to your account.

Tier PlanetScale credit limit
Tier 3 $2,500
Tier 2 $10,000
Tier 1 $50,000

Check your program tier to confirm the limit that applies to your account.

To start using it, open the Hyperdrive section of the dashboard. The Create PlanetScale button appears there for eligible accounts.

Go to Create a PlanetScale database ↗

For database setup and how Hyperdrive connects Workers to PlanetScale, refer to PlanetScale Postgres and MySQL with Hyperdrive.

Custom CA and mTLS certificates for Hyperdrive over Workers VPC

Hyperdrive can now connect to private databases over Workers VPC with a custom certificate authority (CA) and client certificate authentication (mutual TLS, or mTLS). This covers PostgreSQL databases and Amazon RDS instances that present certificates signed by a private CA or the RDS CA bundle, and client certificates with RSA or EC P-256 keys.

Previously, a database that required a custom CA or a client certificate meant routing database traffic over WebSockets through Cloudflare Access and Cloudflare Tunnel. That approach added an Access application, a service token, and WebSocket handling to maintain.

Hyperdrive now connects directly to the database over Workers VPC, without Access.

Attach the certificates to the VPC Service, then create a Hyperdrive configuration from that service:

# Upload the private CA and the client certificate
npx wrangler cert upload certificate-authority --ca-cert ./db-ca.pem --name db-ca
npx wrangler cert upload mtls-certificate --cert ./client.pem --key ./client-key.pem --name db-client

# Attach the certificates to the VPC Service for your database
npx wrangler vpc service create my-postgres-db \
  --type tcp \
  --tcp-port 5432 \
  --app-protocol postgresql \
  --hostname <YOUR_DATABASE_HOSTNAME> \
  --tunnel-id <YOUR_TUNNEL_ID> \
  --certificate-authority <CA_CERTIFICATE_ID> \
  --client-certificate <CLIENT_CERTIFICATE_ID> \
  --cert-verification-mode verify_full

# Create the Hyperdrive configuration from the VPC Service
npx wrangler hyperdrive create <YOUR_CONFIG_NAME> \
  --service-id <YOUR_VPC_SERVICE_ID> \
  --database <DATABASE_NAME> \
  --user <DATABASE_USER> \
  --password <DATABASE_PASSWORD> \
  --scheme postgresql

You can use a custom CA, a client certificate, or both. Configure VPC Services with Wrangler 4.164.0 or later, the API, or Terraform provider 5.29.0 or later. Dashboard support for these settings is not yet available.

For setup details and certificate rotation, refer to Connect Hyperdrive to a private database using Workers VPC and Custom CA and client certificates.

A new, consistent language across Images interfaces

Cloudflare Images lets you dynamically compose, manipulate, and encode images using a broad set of optimizations, including resizing, face cropping, and background removal. You can configure optimizations directly in an image URL or programmatically through the cf.image options on a Worker fetch() subrequest.

We redesigned the Images API to provide a consistent, predictable, and intuitive language for developers and AI agents to optimize and manipulate images.

Starting today, the URL interface and cf.image options in Workers use the same parameter structure to apply optimizations. A configurable optimization will have its corresponding settings nested underneath its top-level parameter. Existing syntax in URLs and Worker configurations continues to behave as before.

This release also makes adding borders and overlays available in the URL interface. For example, you can now add a logo or text watermark directly in the URL instead of writing a Worker.

What changed

Optimization parameters are now represented in one of two ways: a compact form for single values and a structured form for related settings.

For example, width=500 requests an exact width, while trim=color asks Images to automatically trim the sides of an image based on its border color. Color-based trimming already supports configuring the border color, tolerance, and retained edge; these settings were previously applied as separate top-level parameters:

Old Images interface with trim settings as separate top-level parameters

Related trim settings appeared alongside independent parameters such as width.

The new structure groups existing settings under a single top-level trim parameter:

New Images interface with trim settings as nested settings

Related trim settings are nested, while width=250 remains a top-level parameter.

In the URL interface, color-based trimming uses the following syntax:

trim=(mode=color,color=red,tolerance=10,keep=5)

The same hierarchy maps directly to cf.image in Workers, where related values that were previously separate or inferred now live in the same object. The structured object uses mode to explicitly identify the selected behavior, while the remaining properties configure that behavior:

trim: {
	mode: "color",
	color: "white",
	tolerance: 10,
	keep: 5
}

A unified parameter model

In Images, optimizations are expressed as parameters, which form the language used to describe every optimization and its intended behavior. width sets the output width, gravity identifies the area that a crop should preserve, and draw defines how image or text overlays are rendered.

Many optimization features began as scalar parameters, pairing a parameter key with one value, such as width=500. As we introduced more capabilities, some parameters gained settings to configure how the optimization worked, such as automatic width selection. We added syntax to support each use case, gradually accumulating different forms for expressing related settings.

A given optimization might follow multiple conventions for different behaviors. For example, color-based trimming relied on separate dotted parameters, while side-based trimming encoded the top, right, bottom, and left settings as four unlabeled positional values:

trim=border,trim.border.color=white,trim.border.tolerance=10,trim.border.keep=5
trim=10;20;10;20

None of these forms were especially challenging in isolation. But together, they required developers to learn several parameter models and agents to infer the relationships between parameters and their settings to produce the correct syntax for each interface.

In the new structure, the Images API distinguishes between two core concepts: A parameter identifies an optimization feature, while a setting configures how that feature behaves.

The compact form continues to assign one value directly to a parameter. When an optimization has configurable behavior, the structured form organizes named settings, lists, and nested groups beneath the parameter, making the relationship between the optimization and its configuration explicit.

The unified parameter model follows these rules:

  • A parameter that needs one value uses the compact form. When the relationship is unambiguous, existing forms like width=500, width=auto (with default settings), and gravity=auto continue to work with the same syntax.
  • A parameter with configurable behavior groups its settings beneath the top-level parameter. When a parameter supports multiple behaviors, the mode setting identifies the selected behavior. Without mode, an expression such as width=(desktop=1400) would require Images to infer auto from the presence of desktop.
  • Across interfaces, groups and lists follow the same nesting rules. A grouped URL parameter maps to an object in cf.image, while a URL list maps to an array. Previously, each feature needed a separate representation and translation rules for each interface. The shared hierarchy eliminates structural drift by using the same vocabulary and relationships in both interfaces.

For developers, a standard form for configuring parameters means fewer special-case conventions to remember and less translation between interfaces. Related settings stay together and a configuration expressed in a URL maps predictably to its cf.image object.

How this affects existing optimizations

The new structure is available for existing optimization features whose settings previously used separate top-level parameters, dotted names, or positional values. It also brings two previously Workers-only capabilities, borders and overlays, to the URL interface.

Use the new structured form to express the following optimizations and capabilities:

Optimization What it does What changed
border Adds a border around the image. Now supported in the URL interface. Previously, this was available only in Workers.
draw Adds an image or text overlay to the source image. Now supported in the transformation URL interface for remote images. Previously, this was available only in Workers.
gravity coordinates Sets the focal point using relative x and y coordinates. Uses named x and y settings in the URL interface.
quality Controls the compression quality of the output image. Groups the default quality and the quality for slow connections beneath quality.
trim Removes pixels around the sides of an image. Uses named settings beneath trim for individual sides, rectangular areas, and border-color detection.
upscale Controls the algorithm used when an image is enlarged. Configured as the upscale setting beneath fit, alongside the selected mode.
width=auto Automatically selects an appropriate output width using available information about the browser and device. Uses named settings under the width parameter for mobile, desktop, and breakpoint widths.

These structured forms are additive. Earlier forms, including gravity=0.5x0.3, dotted trim parameters, and separate top-level parameters continue to work.

Refer to Features for the complete reference of optimization features.

Cloud Connector is now generally available

Cloud Connector is now generally available to all customers.

Cloud Connector lets you route matching requests to public object storage in Cloudflare R2, Amazon S3, Google Cloud Storage, Microsoft Azure Blob Storage, and Oracle Cloud Infrastructure Object Storage. You can create and manage routing rules in the Cloudflare dashboard, using Rules expressions to control which requests are sent to each provider.

Refer to the Cloud Connector documentation to get started.

WAF Emergency Release - 2026-11-18

This release introduces 6 new malformed base64 Cloudflare Managed Ruleset detections for Java4Shell CVE-2026-114938 ↗︎. The rules are being deployed in log mode by default (in lieu of False Positive risk) and can be overridden if needed with supplementary information around the same available through blog post How Cloudflare WAF helps protect against Java4Shell (CVE-2026-114938) ↗︎.

RulesetRule IDLegacy Rule IDDescriptionPrevious ActionNew ActionComments
Cloudflare Managed RulesetN/AJava4Shell - Malformed Base64 Header - CVE-2026-114938 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/AJava4Shell - Malformed Base64 Query Value - CVE-2026-114938 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/AJava4Shell - Malformed Base64 Body Value - CVE-2026-114938 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/AJava4Shell - Malformed Base64 Path Segment - CVE-2026-114938 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/AJava4Shell - Malformed Base64url JWT Component - CVE-2026-114938 ↗︎N/ALogThis is a new detection.
Cloudflare Managed RulesetN/AJava4Shell - Malformed Base64 WebAuthn Value - CVE-2026-114938 ↗︎N/ALogThis is a new detection.

Custom CA and client certificates (mTLS) for Workers VPC

VPC Services can now verify private services that use certificates from your own certificate authority (CA), and authenticate to services that require mutual TLS (mTLS). Previously, Workers VPC only trusted publicly trusted CAs and Cloudflare Origin CA, so connecting to a private CA meant weakening verification with verify_ca or disabled, and origins that required client certificates were unreachable.

Upload your CA certificate and client certificate, then reference them from the TLS settings of a VPC Service:

npx wrangler cert upload certificate-authority --ca-cert ./ca.pem --name private-ca
npx wrangler cert upload mtls-certificate --cert ./client.pem --key ./client-key.pem --name private-client

npx wrangler vpc service create private-db \
  --type tcp \
  --tcp-port 5432 \
  --app-protocol postgresql \
  --hostname db.internal.example \
  --tunnel-id <TUNNEL_ID> \
  --certificate-authority <CA_CERTIFICATE_ID> \
  --client-certificate <CLIENT_CERTIFICATE_ID> \
  --cert-verification-mode verify_full

You can use a custom CA, a client certificate, or both. The settings apply to HTTPS connections from http services and to TLS connections from tcp services, including PostgreSQL databases used with Hyperdrive, such as Amazon RDS instances that use the RDS CA bundle.

You can configure custom CA and client certificates with Wrangler 4.164.0 or later, the API, or Terraform provider 5.29.0 or later.

For details, refer to Custom CA and client certificates.

AI Search can generate file summaries during indexing

AI Search can now summarize your files while it indexes them. For each text document that produces at least three chunks, AI Search uses a large language model (LLM) to generate a summary, topics, and keywords:

  • The summary and keywords are stored in the item's context metadata. A context value you already set is kept.
  • The topics are added to the context header of the file's chunks to help retrieval.

Turn on file summaries in the dashboard, or set indexing_options.file_summary to true when you create or update an instance. File summaries are off by default, and changing the setting triggers a full reindex.

indexing_options is replaced as a whole on every update. Include the current value of every option, or the options you leave out return to their defaults.

The LLM costs are billed to your account. File summaries are not part of the AI Search ingestion allotment. AI Search uses a Workers AI model suited to summarization and calls it through your instance's AI Gateway, or the default gateway if none is set, and the calls are billed as Workers AI usage. Each new or modified file is summarized once, so costs grow with the amount of content you index.

For more information, refer to File summaries.

Cloudflare CLI (cf) 1.0 release candidate

Cloudflare CLI (cf) 1.0.0-rc.1 is a release candidate for cf 1.0. cf has been in beta since late September. This is the release we recommend you try on your own projects, and we want your feedback before 1.0.

To install or update to the latest release:

npm install --global cf@latest

What is stable

We expect the following features to remain stable to cf 1.0:

  • The cloudflare.config.ts configuration format.
  • The Build Output format that cf build writes and cf deploy reads.
  • The commands that create, develop, build, and deploy projects, including cf init, cf dev, cf build, cf deploy, cf previews deploy, and cf workers versions create.
  • The large majority of commands generated from the Cloudflare API and used in the beta.

A small number of less commonly used API commands may still change before 1.0. cf generates these commands from the Cloudflare API schema, so their call path may still change before 1.0.

Upgrade from a beta

This release moves Build Output to version 1, in .cloudflare/output/v1/. cf deploy will automatically create a new build on first run in this location.

The release candidate also moves to caret (^) version ranges for its dependencies, requiring:

  • @cloudflare/vite-plugin 2.0.0-rc.0 or later, for projects that build with Vite.
  • Wrangler 4.163.0 or later, for legacy projects that build with Wrangler.

1.0.0-rc.1 renames a few generated API commands without aliases. For example, cf workers secrets bulk is now cf workers secrets bulk-edit. For the full list, refer to the release notes ↗︎.

If you last updated early in the beta, these earlier changes may also affect you:

  • Prompts, progress, and status messages are written to stderr. Stdout holds only command output, so scripts that read messages from stdout must read stderr instead.
  • Commands generated from camelCase or snake_case API names use kebab-case. For example, cf accounts billing getCredits is now cf accounts billing get-credits.
  • cf previews deploy takes the Preview name as --preview-name instead of a positional argument.
  • Commands run with --local store state in the project's .cloudflare/state directory.
  • cf no longer saves your account choice between runs. Set accountId in cloudflare.config.ts or CLOUDFLARE_ACCOUNT_ID to skip the prompt.
  • Streaming tail consumers move to experimental.streamingTailConsumers in cloudflare.config.ts.
  • Destructive commands run without a terminal exit with an error unless you pass --force.
  • cf supports the latest versions of the Node.js LTS releases and the latest active release of Node.js.

Stream live logs

cf observability tail streams live logs and other observability events as newline-delimited JSON. Select a Worker with --worker, a zone and its subdomains with --zone, or every event in the account with --all. Narrow the results with --filters, as inline JSON or from a file:

cf observability tail --worker my-worker
cf observability tail --worker my-worker --filters @filters.json

Worker events require Workers Logs to be turned on for the Worker. For filter shapes and examples, run cf observability tail --help.

What changed during the beta

Deploy

  • Deploy a directory of static assets with cf deploy ./dist. cf skips the build step and creates an assets-only Worker.
  • cf deploy --no-provision fails when a binding has no resource ID, instead of creating a new resource.
  • cf deploy --force deploys when the Worker was last deployed through the API, or when its remote configuration conflicts with your local Build Output.
  • Redeploying no longer removes secrets that you added outside cloudflare.config.ts.
  • Stack traces from deployed Workers show project paths such as src/index.ts.
  • Python Workers with vendored packages deploy correctly.
  • cf deploy applies SSH settings for Containers managed by Durable Objects.
  • cf deploy --dry-run shows the project setup changes it would make, without making them.
  • Set CLOUDFLARE_DEPENDENCIES_INSTRUMENTATION=false to stop uploading package dependency metadata with each deployment.

Develop

  • cf init creates new Worker projects with Vite+, including checks, formatting, and tests that run in the Workers runtime.
  • cf dev and cf build run the Vite or Vite+ executable installed in your project, including dependencies hoisted in a workspace.
  • cf dev passes arguments such as --port and --host to your framework's development server.
  • cf dev --profile <name> passes the selected authentication profile to your development server.
  • cf dev, cf build, and cf deploy offer to run cf migrate when they find a Wrangler configuration file.
  • On Windows, Ctrl+C exits cleanly when your framework runs through npm, Wrangler and Vite start without spawn EFTYPE errors, and cf d1 migrations create rejects names that Windows cannot store.
  • cf init approves the build scripts that pnpm 11 and later block by default, so installation succeeds.

Sign in and accounts

  • Pass --temporary-account to create a short-lived account without signing in. cf prints a URL you can use to claim the account.
  • cf auth login summarizes the scopes you granted, and cf auth scopes lists them by product.
  • cf auth whoami, also available as cf whoami, prints a short summary in a terminal and full JSON for scripts and agents.
  • cf auth login now uses the system keyring where available.
  • Long-running commands, such as cf observability tail, refresh an expired OAuth login instead of failing.
  • cf auth whoami reports the status of scoped API tokens correctly.
  • Set CLOUDFLARE_API_HEADERS to send custom headers with API requests.

API commands

  • Every command accepts --dry-run to preview what it would do and it redacts sensitive fields in request previews.
  • cf checks --body input before it asks you to sign in or confirm, and rejects array bodies larger than the API allows instead of splitting them across several requests.
  • More request fields are available as flags, including JSON objects and maps, arrays of objects, and file uploads. For example, cf pages deployments create --worker-js @./_worker.js and cf ai to-markdown --files @first.html @second.pdf.
  • Error outputs have been improved with additional messages the API returns, such as the names of missing parameters.
  • cf workers commands accept --worker to select a Worker.
  • --zone finds zones by name in accounts with many zones.
  • cf tunnels quick-start --allowed-mail limits a temporary tunnel to the email addresses and domains you allow.

Scripts and agents

  • cf --version prints only the version number.
  • Shell completions and cf tools list every flag a command accepts.
  • cf cli search has reranked cf deploy and other key commands to improve their ranking for static site and Pages deployment searches.
  • cf docs search finds Cloudflare documentation from the terminal.

Send feedback

To report a bug or tell us what is missing, open an issue in the cf repository ↗︎. For every change in each release, refer to the Cloudflare CLI release notes ↗︎.

Configure a separate WAN gateway for each appliance in a high availability site

You can now give each Cloudflare One Appliance in a high availability (HA) site its own default gateway on a static WAN. Before this change, both appliances in an HA pair shared the same WAN gateway. Now each appliance can connect to its own Internet circuit, such as a different ISP for each appliance, with its own static IP and gateway.

  • In the dashboard, when you configure a static WAN on an HA site, enter the secondary appliance's gateway in the new Secondary gateway address field, next to Primary gateway address.
  • In the API, set static_addressing.secondary_gateway_address when you create or update a WAN. To remove the value with a PATCH request, set it to null.

The field is optional. If you leave it blank, the secondary appliance uses the primary gateway address, which matches how HA sites worked before. You can only use this field on sites in HA mode.

curl --request PATCH \
https://api.cloudflare.com/client/v4/accounts/{account_id}/magic/sites/{site_id}/wans/{wan_id} \
--header "Authorization: Bearer <API_TOKEN>" \
--header "Content-Type: application/json" \
--data '{
	"static_addressing": {
		"address": "203.0.113.2/30",
		"gateway_address": "203.0.113.1",
		"secondary_address": "198.51.100.2/30",
		"secondary_gateway_address": "198.51.100.1"
	}
}'

For details, refer to About high availability configurations.