Changelog

Radar now includes browser detection for Post-quantum (PQ) encryption. The Post-quantum encryption card ↗ now checks whether a user’s browser supports post-quantum encryption. If support is detected, information about the key agreement in use is displayed.

This week’s highlights prioritise an emergency Oracle E-Business Suite RCE rule deployed to block active, high-impact exploitation. Also addressed are high-severity Chaos Mesh controller command-injection flaws that enable unauthenticated in-cluster RCE and potential cluster compromise, plus a form-data multipart boundary issue that permits HTTP Parameter Pollution (HPP). Two new generic SQLi detections were added to catch inline-comment obfuscation and information disclosure techniques.

Key Findings

• New emergency rule released for Oracle E-Business Suite (CVE-2025-61882) addressing an actively exploited remote code execution vulnerability in core business application modules. Immediate mitigation deployed to protect enterprise workloads.

• Chaos Mesh (CVE-2025-59358,CVE-2025-59359,CVE-2025-59360,CVE-2025-59361): A GraphQL debug endpoint on the Chaos Controller Manager is exposed without authentication; several controller mutations (cleanTcs, killProcesses, cleanIptables) are vulnerable to OS command injection.

• Form-Data (CVE-2025-7783): Attackers who can observe Math.random() outputs and control request fields in form-data may exploit this flaw to perform HTTP parameter pollution, leading to request tampering or data manipulation.

• Two new generic SQLi detections added to enhance baseline coverage against inline-comment obfuscation and information disclosure attempts.

Impact

• CVE-2025-61882 — Oracle E-Business Suite remote code execution (emergency detection): attacker-controlled input can yield full system compromise, data exfiltration, and operational outage; immediate blocking enforced.

• CVE-2025-59358 / CVE-2025-59359 / CVE-2025-59360 / CVE-2025-59361 — Unauthenticated command-injection in Chaos Mesh controllers allowing remote code execution, cluster compromise, and service disruption (high availability risk).

• CVE-2025-7783 — Predictable multipart boundaries in form-data enabling HTTP Parameter Pollution; results include request tampering, parameter overwrite, and downstream data integrity loss.

Managed Ruleset Updated

This update introduces 21 new detections in the Cloudflare Managed Ruleset (all currently set to Disabled mode to preserve remediation logic and allow quick activation if needed). The rules cover a broad spectrum of threats - SQL injection techniques, command and code injection, information disclosure of common files, URL anomalies, and cross-site scripting.

You can now enable Cloudflare Access for your workers.dev and Preview URLs in a single click.

Access allows you to limit access to your Workers to specific users or groups. You can limit access to yourself, your teammates, your organization, or anyone else you specify in your Access policy.

To enable Cloudflare Access:

1. In the Cloudflare dashboard, go to the Workers & Pages page.

   Go to Workers & Pages ↗

2. In Overview, select your Worker.

3. Go to Settings > Domains & Routes.

4. For workers.dev or Preview URLs, click Enable Cloudflare Access.

5. Optionally, to configure the Access application, click Manage Cloudflare Access. There, you can change the email addresses you want to authorize. View Access policies to learn about configuring alternate rules.

To fully secure your application, it is important that you validate the JWT that Cloudflare Access adds to the Cf-Access-Jwt-Assertion header on the incoming request.

The following code will validate the JWT using the jose NPM package ↗:

import { jwtVerify, createRemoteJWKSet } from "jose";

export default {
	async fetch(request, env, ctx) {
		// Verify the POLICY_AUD environment variable is set
		if (!env.POLICY_AUD) {
			return new Response("Missing required audience", {
				status: 403,
				headers: { "Content-Type": "text/plain" },
			});
		}

		// Get the JWT from the request headers
		const token = request.headers.get("cf-access-jwt-assertion");

		// Check if token exists
		if (!token) {
			return new Response("Missing required CF Access JWT", {
				status: 403,
				headers: { "Content-Type": "text/plain" },
			});
		}

		try {
			// Create JWKS from your team domain
			const JWKS = createRemoteJWKSet(
				new URL(`${env.TEAM_DOMAIN}/cdn-cgi/access/certs`),
			);

			// Verify the JWT
			const { payload } = await jwtVerify(token, JWKS, {
				issuer: env.TEAM_DOMAIN,
				audience: env.POLICY_AUD,
			});

			// Token is valid, proceed with your application logic
			return new Response(`Hello ${payload.email || "authenticated user"}!`, {
				headers: { "Content-Type": "text/plain" },
			});
		} catch (error) {
			// Token verification failed
			return new Response(`Invalid token: ${error.message}`, {
				status: 403,
				headers: { "Content-Type": "text/plain" },
			});
		}
	},
};

Add these environment variables to your Worker:

• POLICY_AUD: Your application's AUD tag
• TEAM_DOMAIN: https://<your-team-name>.cloudflareaccess.com

Both of these appear in the modal that appears when you enable Cloudflare Access.

You can set these variables by adding them to your Worker's Wrangler configuration file, or via the Cloudflare dashboard under Workers & Pages > your-worker > Settings > Environment Variables.

Fine-grained permissions for Access Applications, Identity Providers (IdPs), and Targets is now available in Public Beta. This expands our RBAC model beyond account & zone-scoped roles, enabling administrators to grant permissions scoped to individual resources.

• Access Applications ↗: Grant admin permissions to specific Access Applications.
• Identity Providers ↗: Grant admin permissions to individual Identity Providers.
• Targets ↗: Grant admin rights to specific Targets

For more info:

• Get started with Cloudflare Permissioning
• Manage Member Permissioning via the UI & API

Deepgram's newest Flux model @cf/deepgram/flux is now available on Workers AI, hosted directly on Cloudflare's infrastructure. We're excited to be a launch partner with Deepgram and offer their new Speech Recognition model built specifically for enabling voice agents. Check out Deepgram's blog ↗ for more details on the release.

The Flux model can be used in conjunction with Deepgram's speech-to-text model @cf/deepgram/nova-3 and text-to-speech model @cf/deepgram/aura-1 to build end-to-end voice agents. Having Deepgram on Workers AI takes advantage of our edge GPU infrastructure, for ultra low latency voice AI applications.

For the month of October 2025, Deepgram's Flux model will be free to use on Workers AI. Official pricing will be announced soon and charged after the promotional pricing period ends on October 31, 2025. Check out the model page for pricing details in the future.

The new Flux model is WebSocket only as it requires live bi-directional streaming in order to recognize speech activity.

1. Create a worker that establishes a websocket connection with @cf/deepgram/flux

export default {
  async fetch(request, env, ctx): Promise<Response> {
    const resp = await env.AI.run("@cf/deepgram/flux", {
      encoding: "linear16",
      sample_rate: "16000"
    }, {
      websocket: true
    });
    return resp;
  },
} satisfies ExportedHandler<Env>;

2. Deploy your worker

npx wrangler deploy

3. Write a client script to connect to your worker and start sending random audio bytes to it

const ws = new WebSocket('wss://<your-worker-url.com>');

ws.onopen = () => {
  console.log('Connected to WebSocket');

  // Generate and send random audio bytes
  // You can replace this part with a function
  // that reads from your mic or other audio source
  const audioData = generateRandomAudio();
  ws.send(audioData);
  console.log('Audio data sent');
};

ws.onmessage = (event) => {
  // Transcription will be received here
  // Add your custom logic to parse the data
  console.log('Received:', event.data);
};

ws.onerror = (error) => {
  console.error('WebSocket error:', error);
};

ws.onclose = () => {
  console.log('WebSocket closed');
};

// Generate random audio data (1 second of noise at 44.1kHz, mono)
function generateRandomAudio() {
  const sampleRate = 44100;
  const duration = 1;
  const numSamples = sampleRate * duration;
  const buffer = new ArrayBuffer(numSamples * 2);
  const view = new Int16Array(buffer);

  for (let i = 0; i < numSamples; i++) {
    view[i] = Math.floor(Math.random() * 65536 - 32768);
  }

  return buffer;
}

You can now perform more powerful queries directly in Workers Analytics Engine ↗ with a major expansion of our SQL function library.

Workers Analytics Engine allows you to ingest and store high-cardinality data at scale (such as custom analytics) and query your data through a simple SQL API.

Today, we've expanded Workers Analytics Engine's SQL capabilities with several new functions:

New aggregate functions: ↗

• argMin() - Returns the value associated with the minimum in a group
• argMax() - Returns the value associated with the maximum in a group
• topK() - Returns an array of the most frequent values in a group
• topKWeighted() - Returns an array of the most frequent values in a group using weights
• first_value() - Returns the first value in an ordered set of values within a partition
• last_value() - Returns the last value in an ordered set of values within a partition

New bit functions: ↗

• bitAnd() - Returns the bitwise AND of two expressions
• bitCount() - Returns the number of bits set to one in the binary representation of a number
• bitHammingDistance() - Returns the number of bits that differ between two numbers
• bitNot() - Returns a number with all bits flipped
• bitOr() - Returns the inclusive bitwise OR of two expressions
• bitRotateLeft() - Rotates all bits in a number left by specified positions
• bitRotateRight() - Rotates all bits in a number right by specified positions
• bitShiftLeft() - Shifts all bits in a number left by specified positions
• bitShiftRight() - Shifts all bits in a number right by specified positions
• bitTest() - Returns the value of a specific bit in a number
• bitXor() - Returns the bitwise exclusive-or of two expressions

New mathematical functions: ↗

• abs() - Returns the absolute value of a number
• log() - Computes the natural logarithm of a number
• round() - Rounds a number to a specified number of decimal places
• ceil() - Rounds a number up to the nearest integer
• floor() - Rounds a number down to the nearest integer
• pow() - Returns a number raised to the power of another number

New string functions: ↗

• lowerUTF8() - Converts a string to lowercase using UTF-8 encoding
• upperUTF8() - Converts a string to uppercase using UTF-8 encoding

New encoding functions: ↗

• hex() - Converts a number to its hexadecimal representation
• bin() - Converts a string to its binary representation

New type conversion functions: ↗

• toUInt8() - Converts any numeric expression, or expression resulting in a string representation of a decimal, into an unsigned 8 bit integer

Whether you're building usage-based billing systems, customer analytics dashboards, or other custom analytics, these functions let you get the most out of your data. Get started with Workers Analytics Engine and explore all available functions in our SQL reference documentation.

The GraphQL Analytics API now supports confidence intervals for sum and count fields on adaptive (sampled) datasets. Confidence intervals provide a statistical range around sampled results, helping verify accuracy and quantify uncertainty.

• Supported datasets: Adaptive (sampled) datasets only.
• Supported fields: All sum and count fields.
• Usage: The confidence level must be provided as a decimal between 0 and 1 (e.g. 0.90, 0.95, 0.99).
• Default: If no confidence level is specified, no intervals are returned.

For examples and more details, see the GraphQL Analytics API documentation.

New instance types provide up to 4 vCPU, 12 GiB of memory, and 20 GB of disk per container instance.

The dev and standard instance types are preserved for backward compatibility and are aliases for lite and standard-1, respectively. The standard-1 instance type now provides up to 8 GB of disk instead of only 4 GB.

See the getting started guide to deploy your first Container, and the limits documentation for more details on the available instance types and limits.

You can now enhance your security posture by blocking additional application installer and disk image file types with Cloudflare Gateway. Preventing the download of unauthorized software packages is a critical step in securing endpoints from malware and unwanted applications.

We have expanded Gateway's file type controls to include:

• Apple Disk Image (dmg)
• Microsoft Software Installer (msix, appx)
• Apple Software Package (pkg)

You can find these new options within the Upload File Types and Download File Types selectors when creating or editing an HTTP policy. The file types are categorized as follows:

• System: Apple Disk Image (dmg)
• Executable: Microsoft Software Installer (msix), Microsoft Software Installer (appx), Apple Software Package (pkg)

To ensure these file types are blocked effectively, please note the following behaviors:

• DMG: Due to their file structure, DMG files are blocked at the very end of the transfer. A user's download may appear to progress but will fail at the last moment, preventing the browser from saving the file.
• MSIX: To comprehensively block Microsoft Software Installers, you should also include the file type Unscannable. MSIX files larger than 100 MB are identified as Unscannable ZIP files during inspection.

To get started, go to your HTTP policies in Zero Trust. For a full list of file types, refer to supported file types.

Users can now specify that they want to retrieve Cloudflare documentation as markdown rather than the previous HTML default. This can significantly reduce token consumption when used alongside Large Language Model (LLM) tools.

curl https://developers.cloudflare.com/workers/ -H 'Accept: text/markdown'  -v

If you maintain your own site and want to adopt this practice using Cloudflare Workers for your own users you can follow the example here ↗.

A new GA release for the Windows WARP client is now available on the stable releases downloads page.

This release contains minor fixes and improvements.

Changes and improvements

• MASQUE is now the default tunnel protocol for all new WARP device profiles.
• Improvement to limit idle connections in Gateway with DoH mode to avoid unnecessary resource usage that can lead to DoH requests not resolving.
• Improvement to maintain TCP connections to reduce interruptions in long-lived connections such as RDP or SSH.
• Improvements to maintain Global WARP override settings when switching between organizations.
• Improvements to maintain client connectivity during network changes.

Known issues

• For Windows 11 24H2 users, Microsoft has confirmed a regression that may lead to performance issues like mouse lag, audio cracking, or other slowdowns. Cloudflare recommends users experiencing these issues upgrade to a minimum Windows 11 24H2 KB5062553 or higher for resolution.

• Devices using WARP client 2025.4.929.0 and up may experience Local Domain Fallback failures if a fallback server has not been configured. To configure a fallback server, refer to Route traffic to fallback server.

• Devices with KB5055523 installed may receive a warning about Win32/ClickFix.ABA being present in the installer. To resolve this false positive, update Microsoft Security Intelligence to version 1.429.19.0 or later.

• DNS resolution may be broken when the following conditions are all true:

  • WARP is in Secure Web Gateway without DNS filtering (tunnel-only) mode.
  • A custom DNS server address is configured on the primary network adapter.
  • The custom DNS server address on the primary network adapter is changed while WARP is connected.

  To work around this issue, reconnect the WARP client by toggling off and back on.

A new GA release for the macOS WARP client is now available on the stable releases downloads page.

This release contains minor fixes and improvements.

Changes and improvements

• Fixed a bug preventing the warp-diag captive-portal command from running successfully due to the client not parsing SSID on macOS.
• Improvements to maintain Global WARP override settings when switching between organizations.
• MASQUE is now the default tunnel protocol for all new WARP device profiles.
• Improvement to limit idle connections in Gateway with DoH mode to avoid unnecessary resource usage that can lead to DoH requests not resolving.
• Improvements to maintain client connectivity during network changes.
• The WARP client now supports macOS Tahoe (version 26.0).

Known issues

• macOS Sequoia: Due to changes Apple introduced in macOS 15.0.x, the WARP client may not behave as expected. Cloudflare recommends the use of macOS 15.4 or later.

• Devices using WARP client 2025.4.929.0 and up may experience Local Domain Fallback failures if a fallback server has not been configured. To configure a fallback server, refer to Route traffic to fallback server.

A new GA release for the Linux WARP client is now available on the stable releases downloads page.

This release contains minor fixes and improvements including an updated public key for Linux packages. The public key must be updated if it was installed before September 12, 2025 to ensure the repository remains functional after December 4, 2025. Instructions to make this update are available at pkg.cloudflareclient.com.

Changes and improvements

• MASQUE is now the default tunnel protocol for all new WARP device profiles.
• Improvement to limit idle connections in Gateway with DoH mode to avoid unnecessary resource usage that can lead to DoH requests not resolving.
• Improvements to maintain Global WARP override settings when switching between organizations.
• Improvements to maintain client connectivity during network changes.

Known issues

• Devices using WARP client 2025.4.929.0 and up may experience Local Domain Fallback failures if a fallback server has not been configured. To configure a fallback server, refer to Route traffic to fallback server.

Gateway users can now apply granular controls to their file sharing and AI chat applications through HTTP policies.

The new feature offers two methods of controlling SaaS applications:

• Application Controls are curated groupings of Operations which provide an easy way for users to achieve a specific outcome. Application Controls may include Upload, Download, Prompt, Voice, and Share depending on the application.
• Operations are controls aligned to the most granular action a user can take. This provides a fine-grained approach to enforcing policy and generally aligns to the SaaS providers API specifications in naming and function.

Get started using Application Granular Controls and refer to the list of supported applications.

Radar now introduces Regional Data, providing traffic insights that bring a more localized perspective to the traffic trends shown on Radar.

The following API endpoints are now available:

• Get Geolocation - Retrieves geolocation by geoId.
• List Geolocations - Lists geolocations.
• NetFlows Summary By Dimension - Retrieves NetFlows summary by dimension.

All summary and timeseries_groups endpoints in HTTP and NetFlows now include an adm1 dimension for grouping data by first level administrative division (for example, state, province, etc.)

A new filter geoId was also added to all endpoints in HTTP and NetFlows, allowing filtering by a specific administrative division.

Check out the new Regional traffic insights on a country specific traffic page new Radar page ↗.

This week highlights four important vendor- and component-specific issues: an authentication bypass in SimpleHelp (CVE-2024-57727), an information-disclosure flaw in Flowise Cloud (CVE-2025-58434), an SSRF in the WordPress plugin Ditty (CVE-2025-8085), and a directory-traversal bug in Vite (CVE-2025-30208). These are paired with improvements to our generic detection coverage (SQLi, SSRF) to raise the baseline and reduce noisy gaps.

Key Findings

• SimpleHelp (CVE-2024-57727): Authentication bypass in SimpleHelp that can allow unauthorized access to management interfaces or sessions.

• Flowise Cloud (CVE-2025-58434): Information-disclosure vulnerability in Flowise Cloud that may expose sensitive configuration or user data to unauthenticated or low-privileged actors.

• WordPress:Plugin: Ditty (CVE-2025-8085): SSRF in the Ditty WordPress plugin enabling server-side requests that could reach internal services or cloud metadata endpoints.

• Vite (CVE-2025-30208): Directory-traversal vulnerability in Vite allowing access to filesystem paths outside the intended web root.

Impact

These vulnerabilities allow attackers to gain access, escalate privileges, or execute actions that were previously unavailable:

• SimpleHelp (CVE-2024-57727): An authentication bypass that can let unauthenticated attackers access management interfaces or hijack sessions — enabling lateral movement, credential theft, or privilege escalation within affected environments.

• Flowise Cloud (CVE-2025-58434): Information-disclosure flaw that can expose sensitive configuration, tokens, or user data; leaked secrets may be chained into account takeover or privileged access to backend services.

• WordPress:Plugin: Ditty (CVE-2025-8085): SSRF that enables server-side requests to internal services or cloud metadata endpoints, potentially allowing attackers to retrieve credentials or reach otherwise inaccessible infrastructure, leading to privilege escalation or cloud resource compromise.

• Vite (CVE-2025-30208): Directory-traversal vulnerability that can expose filesystem contents outside the web root (configuration files, keys, source code), which attackers can use to escalate privileges or further compromise systems.

This week highlights multiple critical Cisco vulnerabilities (CVE-2025-20363, CVE-2025-20333, CVE-2025-20362). This flaw stems from improper input validation in HTTP(S) requests. An authenticated VPN user could send crafted requests to execute code as root, potentially compromising the device.

Key Findings

• Cisco (CVE-2025-20333, CVE-2025-20362, CVE-2025-20363): Multiple vulnerabilities that could allow attackers to exploit unsafe deserialization and input validation flaws. Successful exploitation may result in arbitrary code execution, privilege escalation, or command injection on affected systems.

Impact

Cisco (CVE-2025-20333, CVE-2025-20362, CVE-2025-20363): Exploitation enables attackers to escalate privileges or achieve remote code execution via command injection.

Managed Ruleset Updated

This update introduces 11 new detections in the Cloudflare Managed Ruleset (all currently set to Disabled mode to preserve remediation logic and allow quick activation if needed). The rules cover a broad spectrum of threats - SQL injection techniques, command and code injection, information disclosure of common files, URL anomalies, and cross-site scripting.

The ctx.exports API contains automatically-configured bindings corresponding to your Worker's top-level exports. For each top-level export extending WorkerEntrypoint, ctx.exports will contain a Service Binding by the same name, and for each export extending DurableObject (and for which storage has been configured via a migration), ctx.exports will contain a Durable Object namespace binding. This means you no longer have to configure these bindings explicitly in wrangler.jsonc/wrangler.toml.

Example:

import { WorkerEntrypoint } from "cloudflare:workers";

export class Greeter extends WorkerEntrypoint {
  greet(name) {
    return `Hello, ${name}!`;
  }
}

export default {
  async fetch(request, env, ctx) {
    let greeting = await ctx.exports.Greeter.greet("World")
    return new Response(greeting);
  }
}

At present, you must use the enable_ctx_exports compatibility flag to enable this API, though it will be on by default in the future.

See the API reference for more information.

Today, we're launching the new Cloudflare Pipelines: a streaming data platform that ingests events, transforms them with SQL, and writes to R2 as Apache Iceberg ↗ tables or Parquet files.

Pipelines can receive events via HTTP endpoints or Worker bindings, transform them with SQL, and deliver to R2 with exactly-once guarantees. This makes it easy to build analytics-ready warehouses for server logs, mobile application events, IoT telemetry, or clickstream data without managing streaming infrastructure.

For example, here's a pipeline that ingests clickstream events and filters out bot traffic while extracting domain information:

INSERT into events_table
SELECT
  user_id,
  lower(event) AS event_type,
  to_timestamp_micros(ts_us) AS event_time,
  regexp_match(url, '^https?://([^/]+)')[1]  AS domain,
  url,
  referrer,
  user_agent
FROM events_json
WHERE event = 'page_view'
  AND NOT regexp_like(user_agent, '(?i)bot|spider');

Get started by creating a pipeline in the dashboard or running a single command in Wrangler:

npx wrangler pipelines setup

Check out our getting started guide to learn how to create a pipeline that delivers events to an Iceberg table you can query with R2 SQL. Read more about today's announcement in our blog post ↗.

Today, we're launching the open beta for R2 SQL: A serverless, distributed query engine that can efficiently analyze petabytes of data in Apache Iceberg ↗ tables managed by R2 Data Catalog.

R2 SQL is ideal for exploring analytical and time-series data stored in R2, such as logs, events from Pipelines, or clickstream and user behavior data.

If you already have a table in R2 Data Catalog, running queries is as simple as:

npx wrangler r2 sql query YOUR_WAREHOUSE "
SELECT
    user_id,
    event_type,
    value
FROM events.user_events
WHERE event_type = 'CHANGELOG' or event_type = 'BLOG'
  AND __ingest_ts > '2025-09-24T00:00:00Z'
ORDER BY __ingest_ts DESC
LIMIT 100"

To get started with R2 SQL, check out our getting started guide or learn more about supported features in the SQL reference. For a technical deep dive into how we built R2 SQL, read our blog post ↗.

We’re shipping three updates to Browser Rendering:

• Playwright support is now Generally Available and synced with Playwright v1.55 ↗, giving you a stable foundation for critical automation and AI-agent workflows.
• We’re also adding Stagehand support (Beta) so you can combine code with natural language instructions to build more resilient automations.
• Finally, we’ve tripled limits for paid plans across both the REST API and Browser Sessions to help you scale.

To get started with Stagehand, refer to the Stagehand example that uses Stagehand and Workers AI to search for a movie on this example movie directory ↗, extract its details using natural language (title, year, rating, duration, and genre), and return the information along with a screenshot of the webpage.

const stagehand = new Stagehand({
	env: "LOCAL",
	localBrowserLaunchOptions: { cdpUrl: endpointURLString(env.BROWSER) },
	llmClient: new WorkersAIClient(env.AI),
	verbose: 1,
});

await stagehand.init();
const page = stagehand.page;

await page.goto("https://demo.playwright.dev/movies");

// if search is a multi-step action, stagehand will return an array of actions it needs to act on
const actions = await page.observe('Search for "Furiosa"');
for (const action of actions) await page.act(action);

await page.act("Click the search result");

// normal playwright functions work as expected
await page.waitForSelector(".info-wrapper .cast");

let movieInfo = await page.extract({
	instruction: "Extract movie information",
	schema: z.object({
		title: z.string(),
		year: z.number(),
		rating: z.number(),
		genres: z.array(z.string()),
		duration: z.number().describe("Duration in minutes"),
	}),
});

await stagehand.close();

AutoRAG is now AI Search! The new name marks a new and bigger mission: to make world-class search infrastructure available to every developer and business.

With AI Search you can now use models from different providers like OpenAI and Anthropic. By attaching your provider keys to the AI Gateway linked to your AI Search instance, you can use many more models for both embedding and inference.

To use AI Search with other model providers:

1. Add provider keys to AI Gateway
   1. Go to AI > AI Gateway in the dashboard.
   2. Select or create an AI gateway.
   3. In Provider Keys, choose your provider, click Add, and enter the key.
2. Connect a gateway to AI Search: When creating a new AI Search, select the AI Gateway with your provider keys. For an existing AI Search, go to Settings and switch to a gateway that has your keys under Resources.
3. Select models: Embedding models are only available to be changed when creating a new AI Search. Generation model can be selected when creating a new AI Search and can be changed at any time in Settings.

Once configured, your AI Search instance will be able to reference models available through your AI Gateway when making a /ai-search request:

export default {
  async fetch(request, env) {
    
    // Query your AI Search instance with a natural language question to an OpenAI model
    const result = await env.AI.autorag("my-ai-search").aiSearch({
      query: "What's new for Cloudflare Birthday Week?",
      model: "openai/gpt-5"
    });

    // Return only the generated answer as plain text
    return new Response(result.response, {
      headers: { "Content-Type": "text/plain" },
    });
  },
};

In the coming weeks we will also roll out updates to align the APIs with the new name. The existing APIs will continue to be supported for the time being. Stay tuned to the AI Search Changelog and Discord ↗ for more updates!

You can now run more Containers concurrently with higher limits on CPU, memory, and disk.

You can now run 1000 instances of the dev instance type, 400 instances of basic, or 100 instances of standard concurrently.

This opens up new possibilities for running larger-scale workloads on Containers.

See the getting started guide to deploy your first Container, and the limits documentation for more details on the available instance types and limits.