September 2026 Summaries

Cloudflare argues that AI agents have become a major second audience for the Internet, with AI-agent requests on its network rising more than 1,700% over the past year and nonhuman traffic now exceeding half of overall traffic, while websites face declining human referrals, advertising exposure, and revenue despite increasing infrastructure costs. It distinguishes agents acting on users’ behalf from training crawlers, noting that agents may represent potential customers and therefore require more nuanced controls than blanket bot blocking. Cloudflare describes tools for identifying verified bots, analyzing their activity, and separately managing search, agent, and training access, including mechanisms intended to keep sites searchable while restricting AI training use. To support new business models, it is developing Pay Per Use for compensating publishers when AI systems use content and a Monetization Gateway that enables sites to charge agents directly through HTTP 402 and the x402 protocol for data, APIs, and other services. The company is also pursuing efficiency measures such as agent-readable Markdown pages, WebMCP actions, crawler bandwidth reporting, and research with OpenAI on more efficient discovery and indexing, while advocating open standards and interoperable payment, identity, and access systems rather than centralized control of the emerging agent-driven web.

Cloudflare has introduced Pay Per Use in beta, a system intended to let publishers earn revenue when AI products use their content rather than merely crawl it. AI companies define qualifying uses, such as displaying a cited excerpt or using a review to inform a recommendation, set a price, identify their verified crawlers, and report each usage event through a single API; publishers can review and accept or decline these offers through Cloudflare’s dashboard without separate integrations or contracts. Cloudflare validates that reported uses correspond to enrolled publishers, aggregates usage, bills buyers, and distributes monthly payments, while providing publishers with reporting on usage and estimated earnings by buyer and domain. The service complements Cloudflare’s earlier Pay Per Crawl model by focusing on downstream value rather than access, and it aims to offer buyers efficient access to current, paywalled, or specialized content. During the beta, Cloudflare plans to evaluate pricing, reporting, and settlement with participating buyers and publishers, while developing features such as use-context reporting and flexible pricing by content type or use. Pay Per Use is presented as part of a broader monetization framework for the agentic web, alongside Monetization Gateway for APIs and tools that charge agents per request.

Cloudflare has introduced a new durable_object scheduling policy for Containers, aimed at agent workloads that require on-demand, configurable, persistent Linux sandboxes. The update lets application code running in Durable Objects select a Container’s image and compute instance type at runtime, rather than requiring separate preconfigured deployments, while also enabling code-managed rollout strategies. Cloudflare reports that a redesigned scheduling path and runtime reduce median Container startup time from 4.049 seconds to 648 milliseconds in an independent benchmark, with preliminary internal tests creating 100,000 Containers in 5.387 seconds across six locations. A new prepared Debian Trixie system image is intended to reduce setup overhead, and public-beta filesystem snapshots allow workspaces to be saved, restored, and reused as consistent baselines for tasks such as coding sessions, evaluations, and reinforcement learning. The company positions Durable Objects as persistent controllers for Container identity, state, lifecycle, networking, credentials, and supervision, with new native ctx.container APIs replacing much of the prior wrapper-based model. New functionality is available only through the native Durable Object approach, while the existing Container and legacy Sandbox classes will be maintained without updates through December 31, 2026, and Sandbox SDK 1.0 is being refocused as a collection of utilities for developers’ own Durable Object classes.

Cloudflare has launched its Monetization Gateway in closed beta, enabling eligible U.S.-based domain owners to charge AI agents per use for websites, APIs, MCP tools, and datasets through HTTP 402 Payment Required responses and x402-based payments settled in USDC on the Base blockchain. Designed for agent consumption patterns that favor low-cost, fast, automated transactions over subscriptions and prepaid credits, the service lets sellers configure access rules and fixed, variable, or origin-controlled pricing while Cloudflare manages payment verification, settlement, retries, analytics, and protocol updates via Coinbase’s x402 Facilitator. Initial production use cases include Cloudflare AI Gateway, where agents can pay per inference request; Ceramic.ai, which sells individual agent-oriented web searches; Stocktwits, which provides on-demand market sentiment and activity signals; and API2PDF, which charges agents for PDF-generation requests based on actual compute and bandwidth use. Cloudflare plans to add agent discovery, transaction logs, further payment rails, and identity capabilities as it expands access to additional geographies.

Cloudflare has expanded AI Gateway User Insights to help organizations understand AI usage beyond model names, token counts, and request volumes by classifying conversations into task categories such as coding, research, writing, summarization, and data analysis. The update includes an “overkill” view that identifies cases where a high-capability model may be handling work that could potentially be completed by a faster or less expensive option, while allowing teams to investigate related users, agents, applications, costs, latency, token use, and conversation turns before changing workflows. These signals also support Potential Savings and an Auto Router, available in closed beta, which can select models based on task complexity, conversation patterns, model fit, and cost rather than automatically choosing the cheapest option. Classification is performed asynchronously through a Cloudflare Worker using eligible AI Gateway logs, produces derived categories and aggregate reporting rather than exposing raw prompts, and may lag traffic by about one day. User Insights can associate activity with users, teams, and applications through AI Gateway metadata and Cloudflare Access, and the capabilities are available free to AI Gateway users.

Cloudflare has introduced Issues, an open-beta error-monitoring feature for Workers that automatically groups recurring exceptions, HTTP 5xx responses, error logs, and other failures into consolidated incidents with diagnostic context such as stack traces, logs, traces, request data, Worker versions, and frequency trends. Enabled through a single configuration setting without additional SDKs or instrumentation, it can also detect conditions including runaway alarms and excessive loop-generated logging. Developers can enrich issue data with user, account, and session identifiers through the Workers runtime’s OpenTelemetry API, helping identify affected application contexts. Through Cloudflare Automations, Issues can send incident summaries to coding agents such as Claude Code, Cursor, and Devin, generic webhooks, or team communication and incident-management tools, while Cloudflare MCP can provide agents with access to related telemetry for deeper investigation and pull-request generation. Cloudflare reports that internally testing Issues on its Workflows platform surfaced and helped resolve a migration retry-loop bug and an incomplete deletion process caused by subrequest limits, and users can begin by enabling Issues in wrangler.jsonc and configuring an automation destination.

Cloudflare has redesigned its Registrar experience to simplify domain discovery and purchasing while maintaining at-cost pricing and avoiding unnecessary upsells. Its new browser search covers more than 420 domain extensions, displays exact search terms across all supported extensions, includes registered and available domains, and offers real-time sorting, filtering, transparent registration and renewal prices, and selected first-year discounts. Cloudflare has also expanded its Registrar API with a sandbox, extension details, and programmatic transfers, making these features accessible through Cloudflare MCP and the cf CLI for AI agents and developers. The search system combines prepared registry datasets, DNS data, cached results, and live registry queries to balance speed, accuracy, and upstream request limits, progressively updating individual results as stronger availability evidence becomes available. Built using Cloudflare Workers, Durable Objects, Workers KV, and WebSockets, the platform coordinates large numbers of extension checks efficiently while delivering dynamic results in the browser.

Cloudflare reports that its Impact programs have reached $100 million in donated services, supporting public-interest organizations such as journalists, civil society groups, election bodies, schools, and governments with cybersecurity tools. Originating from its free-service model, the company launched Project Galileo in 2014 for vulnerable organizations facing sophisticated attacks; it now protects more than 3,500 domains in over 120 countries and blocked 38.5 billion attacks in 2025. Other initiatives include the Athenian Project for election infrastructure, Cloudflare for Campaigns, and protections for critical services including schools, health clinics, Ukrainian government entities, and community networks. Cloudflare’s 2026 civil society threat report found that Project Galileo participants faced website vulnerability exploit attempts more than seven times as often as average customers. The company is also expanding its efforts beyond defense, including AI crawler controls for local news organizations, startup credits for nonprofits, developer-platform projects for human rights groups, and a Barcelona hackathon where journalists and technologists developed tools for AI bias research, newsroom workflows, content verification, and information integrity.

Cloudflare has launched Auto Router in public beta through AI Gateway, allowing organizations to send requests to the cloudflare/auto model endpoint and automatically select an appropriate AI model based on task requirements, expected quality, and cost. Designed to reduce reliance on manual model selection and avoid using expensive frontier models for simpler work, the service reportedly achieved up to 30% savings in Cloudflare’s internal use and delivered an 86.6% success rate on an internal knowledge-work benchmark at lower cost than OpenAI GPT-6 Sol and Anthropic Claude Opus 5.5. The router filters eligible models according to compatibility, policies, billing, health, and access controls, then uses an edge-hosted classifier to assess task category, complexity, ambiguity, stakes, and contextual dependence before ranking candidates through a quality-versus-cost utility calculation. It also considers cache and context-rewriting costs in longer agentic sessions, making model switches less likely when existing session context is valuable. Cloudflare plans to expand supported models and add capabilities such as zero-data-retention filtering, provider-capacity awareness, configurable reasoning levels, and support for additional APIs, while the service remains free during its beta period.

Cloudflare has announced plans to become a publicly trusted certificate authority, applying to the root programs operated by Chrome, Apple, Microsoft, and Mozilla while acquiring an established GlobalSign root certificate to provide broad compatibility, including for older devices. The company aims to offer free, automated, ACME-based certificate issuance as an additional source of capacity and redundancy alongside existing providers such as Let’s Encrypt, responding to growing certificate demand and risks associated with dependence on a small number of CAs. Cloudflare says its CA will require automated renewal support through ACME Renewal Information, use mechanisms intended to limit the impact of revocation or compliance incidents, and provide operational transparency through reproducible software builds, hardware attestations, and public health reporting. It also plans to issue post-quantum Merkle Tree Certificates in production beginning in early 2027, supporting both conventional and post-quantum certificate formats during a gradual industry transition, while first testing its new CA infrastructure across Cloudflare’s own operations.

Cloudflare aims to achieve full post-quantum cryptographic readiness by 2029, extending its existing post-quantum TLS encryption deployments to authentication, legacy connections, and other public-key systems. To manage the scale and complexity of identifying cryptography across numerous repositories, configurations, dependencies, and protocols, it is developing CryptoLabe, an internal AI-assisted tool that discovers cryptographic uses, analyzes their runtime context, classifies classical and post-quantum implementations, tracks migration metrics, and identifies shared prerequisites or difficult cases. Built on Cloudflare’s Developer Platform using Workers, Durable Objects, Workflows, R2, D1, AI Gateway, Workers AI, and isolated Sandboxes, CryptoLabe uses staged scans and engineering review to produce actionable reports for product managers and developers while avoiding unsupported conclusions. Cloudflare emphasizes that post-quantum migration depends on standards, libraries, vendors, clients, and other ecosystem participants, and recommends that other organizations prioritize their most sensitive or exposed systems, apply broad protective controls where available, validate findings with system owners, and address high-impact upgrades and common blockers before attempting exhaustive cryptographic inventories.

Web Public Key Infrastructure relies on browsers, certificate authorities, and certificate transparency logs to verify website identities, but the expected transition to post-quantum cryptography by 2029 poses major scaling challenges because post-quantum signatures are far larger than current signatures. Merkle Tree Certificates (MTCs), an emerging IETF design, aim to address this by integrating certificate issuance with append-only transparency logs, allowing certificate authorities to sign batches of certificates through Merkle-tree checkpoints while clients verify compact inclusion proofs. Cloudflare plans to support free MTC issuance through its new certificate authority, with a target of inclusion in Chrome’s Quantum-resistant Root Store in early 2027, while also operating transparency-log mirrors and requiring independent cosigning. In an experiment with Chrome Beta involving billions of certificate deliveries, landmark-relative MTCs reduced handshake data and achieved median performance 9% faster than traditional certificate chains, although standalone fallbacks remain necessary for offline or newly installed clients. Wider deployment will still require independent monitoring, diverse certificate authorities and cosigners, browser support, and rigorous evaluation of Cloudflare’s CA by Chrome’s root program.

Cloudflare argues that AI agents are making cyberattacks faster, more persistent, and more capable of chaining vulnerabilities, credentials, and permissions, citing a reported OpenAI and Hugging Face compromise in which agents allegedly progressed from code execution to broad administrative access in under 13 hours after months of precursor activity. It contends that isolated security tools and alerts are insufficient because AI-enabled attacks can bypass network restrictions, use valid credentials, mutate payloads, and exploit business logic, while AI-assisted software development and undiscovered software dependencies expand organizational risk. Cloudflare proposes a connected four-stage application-security framework covering risk discovery and prioritization, governance of human and agent access, runtime protection, and investigation and response, supported by its network-scale threat intelligence, local application context, and inline enforcement. Announced and developing capabilities include LLM-assisted code vulnerability discovery, continuous agentic penetration testing, Botbase identity registration, behavioral analysis through Precursor, Application Profiles for positive security models, model-enhanced WAF rules, machine-learning attack detection, AI application guardrails, expanded threat intelligence access, and agent-supported security operations that correlate events across external application and internal corporate traffic.

Cloudflare has launched Application Profiles, extending its API Schema Learning and Validation capabilities to web applications to support positive security policies that identify and optionally block HTTP requests deviating from learned, expected structures. The feature analyzes selected operations’ paths, query parameters, headers, cookies, and JSON or form-encoded bodies to learn data types and constraints such as UUID formats, numeric ranges, enumerations, string lengths, and permitted characters, helping reduce exposure to attacks including SQL injection, cross-site scripting, and remote code execution without requiring known attack signatures. Profiles are learned from successful traffic and continuously updated, while validation initially adds metadata rather than automatically blocking requests, allowing customers to investigate violations through Security Analytics and create targeted Security Rules when ready. Cloudflare recommends observation before enforcement because legitimate releases, clients, or uncommon requests may not conform to a profile. The company also plans LLM-powered contextual analysis to identify critical operations and fields, recommend mitigations, and prioritize risks such as reconnaissance, data loss, and business-critical endpoints. The capability is available to API Security customers and is in invitation-only beta for other Enterprise customers, with current limitations including no support for multipart forms, GraphQL, XML, required-parameter enforcement, or blocking solely on new parameters.

Cloudflare has added post-quantum TLS 1.3 visibility to HTTP Traffic Analytics, Log Explorer, and Logpush, enabling customers to monitor the key-exchange algorithms used for individual visitor connections and assess cryptographic readiness by domain. The new telemetry highlights use of the hybrid X25519MLKEM768 algorithm, which combines classical X25519 with post-quantum ML-KEM and is the recommended TLS 1.3 approach for defending against harvest-now-decrypt-later threats, while also identifying traffic that uses classical algorithms, deprecated draft implementations, RSA-based exchanges, or no TLS. Cloudflare reports that roughly 70% of browser traffic reaching its network already uses hybrid ML-KEM, compared with about 15% of Cloudflare-to-origin connections, illustrating a gap in origin adoption. Customers can use the ClientTLSKeyExchangeGroup and OriginTLSKeyExchangeGroup log fields to investigate both visitor-to-Cloudflare and Cloudflare-to-origin encryption, enable TLS 1.3 to allow automatic post-quantum negotiation, or use Cloudflare Tunnel to protect connections to legacy origins. The release is part of Cloudflare’s broader effort to achieve full post-quantum security by 2029, with future plans to expose post-quantum authentication data as certificate and signature technologies become more widely deployed.

Cloudflare evaluated its web application firewall using an AI-driven dynamic testing system that simulated an attacker by iteratively mutating known payloads across 45 scenarios involving XSS, SQL injection, command injection, SSRF, path traversal, Log4j, and log injection in an authorized customer staging environment. The models had no access to source code, WAF rules, or enforcement controls, while Python-based safeguards controlled request replay, allowlisted targets, limits, and evidence collection. Of 1,107 generated attempts, 558 were blocked, and human triage reduced unblocked or ambiguous outputs to 49 WAF-relevant findings, mostly involving command injection and SSRF; these findings were leads for validation rather than evidence of successful exploitation. One SSRF example showed that a trailing-dot representation of a cloud metadata IP address produced a redirect instead of a WAF block, prompting investigation without proving origin access. The exercise led to new detections for obfuscated SSRF hosts and restricted protocols, plus improvements to an existing SSRF rule, and indicated that broader scenario coverage was more useful than simply increasing attempts within a single scenario. Cloudflare recommends defense in depth, including properly configured managed WAF rules and attack scoring, staged testing, positive security controls, additional API and bot protections, and regular software patching because WAF bypasses still require an exploitable application to cause harm.

Cloudflare has collaborated with the IETF to address a quantum-era downgrade vulnerability in IPsec, a widely used network-layer security protocol supporting products such as Cloudflare WAN and Magic Transit. During the transition to post-quantum cryptography, systems must retain classical cryptographic support for compatibility, but attackers could exploit negotiation mechanisms to force quantum-capable endpoints to use weaker classical key exchange and potentially decrypt traffic using a sufficiently capable quantum computer. The identified IPsec flaw stems from IKEv2 authenticating only each party’s outgoing messages rather than the complete handshake transcript, enabling attackers to create inconsistent views of a connection and, in some cases, misbind identities. Cloudflare and the IETF’s IPSECME Working Group developed the IKE_SA_INIT_FULL_TRANSCRIPT_AUTH extension, which makes endpoints authenticate the full handshake transcript and uses unconditional signaling to make removal of the feature detectable. Cloudflare has introduced beta support for the extension through an account-level ipsec_downgrade_protection flag, plans broader availability after testing, and encourages adoption across the IPsec ecosystem as post-quantum migration advances.

Cloudflare has launched Threat Signals, a generally available service that uses agentic AI skills to convert selected open-source threat intelligence RSS reports into actionable, contextualized indicators within each customer’s private Threat Intelligence dataset. The service monitors RSS 2.0, Atom, and RDF feeds, retrieves and cleans article content, summarizes reports, extracts and normalizes indicators of compromise, applies account-specific tags, and preserves links between indicators, threat events, and original sources so analysts can understand their relevance and provenance. These threat events can be investigated through the Threat Events Platform and used to create WAF policies, while all Cloudflare accounts receive free access to one RSS feed, 30 days of private derived data storage, and API and dashboard tools. Higher-tier enterprise customers can add more feeds, storage, proprietary Cloudforce One intelligence, custom AI skills, and custom WAF rules. Cloudflare emphasized that the system was designed to prioritize trusted outputs, familiar tagging vocabularies, transparency about automated tags, and retained source context, and it plans to support additional intelligence formats and ingestion pipelines beyond RSS.

Cloudflare has released Vinext 1.0, an open-source framework designed to make existing Next.js applications portable across platforms including Cloudflare Workers, Netlify, and AWS Lambda while using Vite as its foundation. Originating as a week-long AI-driven experiment, Vinext has evolved over seven months to support production workloads, with more than 99% compatibility in tested core Next.js behavior excluding Cache Components. The release supports both App Router and Pages Router applications, including hybrid projects, React Server Components, Server Actions, API routes, middleware, client navigation, ISR, static exports, observability integrations, and common Next.js ecosystem patterns. Vinext 1.0 also adds build-time prerendering and Cloudflare-based cache warming, which renders selected and high-traffic pages before a deployment receives production traffic to avoid lengthy builds and improve initial response times. While support for Next.js Cache Components remains limited, Cloudflare plans to maintain compatibility through automated monitoring of upstream Next.js changes, nightly test runs, and AI-assisted issue detection and fixes.

Cloudflare has released EmDash 1.0, a free, MIT-licensed open-source CMS built on Astro that combines developer tooling, an editorial admin interface, APIs, CLI access, and an MCP server for AI agents. The stable release follows production use and testing, including migrations by Avulux and Cloudflare’s own blog, which informed features for performance, caching, localization, media management, database migrations, and security at high traffic volumes. EmDash supports a growing ecosystem of themes, e-commerce tools, multisite platforms, and sandboxed plugins distributed through a decentralized registry based on AT Protocol, where publishers retain control of their signed packages and release histories while site owners can inspect requested permissions before installation. Its plugin architecture isolates extensions and grants access to content, media, users, email, or network services only through administrator-approved capabilities, with compatible runtimes on Cloudflare Workers and Node.js. Cloudflare also introduced EmDash Build, an open-source alpha AI site builder that generates full Astro and EmDash websites from prompts, records changes as Git commits, and deploys completed projects through Workers for Platforms.

Cloudflare has introduced Forge, an open-source, Apache 2.0-licensed and pluggable generation pipeline for producing SDKs, CLIs, documentation, libraries, and other API-consumption tools from specifications such as OpenAPI. Built to manage Cloudflare’s API ecosystem of more than 3,500 operations across hundreds of repositories, Forge runs in CI to lint API changes and create per-pull-request preview builds, helping teams identify generation problems before releases. Its chained transformer architecture allows outputs such as TypeScript SDKs to serve as inputs for other outputs, including the cf CLI, Cap’n Web RPC specifications, documentation, MCP servers, validation schemas, and potentially bindings for tools like TanStack Query. Cloudflare says this flexibility also supports handwritten CLI functionality and associated documentation, while enabling future support for formats including AsyncAPI, GraphQL, Protobuf, and Cap’n Proto. Forge already generates requirements for the cf CLI and is expected to expand to Cloudflare’s API documentation and SDKs, while also supporting the company’s effort to introduce API versioning without breaking existing users and clients.

Cloudflare has launched BEACON, an anonymized public dataset of billions of daily real-user web performance measurements from 10,000 large websites, designed to reveal how browser, device, network, geography, and site architecture affect user experience beyond developers’ high-end environments. Available through Google BigQuery and based on RUM Archive standards, the dataset reports Core Web Vitals as full histograms, allowing researchers to examine slow experiences in the long tail rather than relying only on 75th-percentile results. Early findings show performance differences among browser engines and industry categories, while LCP and INP sub-metrics indicate that delayed resource discovery, rendering blocks, JavaScript execution, and complex layout work often matter more than download time itself. BEACON also shows that client-side soft navigations are substantially faster than hard navigations, though their benefits must be weighed against potentially slower initial landing pages. By combining the data with sources such as Cloudflare’s Internet Quality Index and World Bank economic indicators, researchers can investigate how network quality and economic conditions relate to web performance, including evidence that lower-bandwidth regions may receive smaller page transfers. Cloudflare says privacy protections remove site identifiers, normalize traffic across included sites, and aggregate records while excluding small groups, with plans to add metrics and incorporate the findings into Cloudflare Radar’s Web Performance section.

Four months after joining Cloudflare, VoidZero reports maintaining its commitment to keep Vite, Vitest, Rolldown, Oxc, and Vite+ open source, vendor-neutral, and community-driven while releasing more than 80 updates and closing over 1,200 issues. Key performance developments include the Rust-based Oxc React Compiler, said to compile React applications 10 times faster than the Babel implementation; Vitest 5, which can reduce test times by up to 50%; stable type-aware tsgolint, reportedly 12 to 18 times faster than ESLint in large TypeScript projects; and Rust-based Oxfmt formatters that are claimed to run seven times faster than Prettier while retaining compatible output. The organization is also advancing Bundled Dev, a Vite development mode using production bundling for improved performance in large applications, and has released Vite+ 1.0 as an integrated toolchain including Vite, Vitest, Rolldown, Oxlint, Oxfmt, and task caching. VoidZero frames faster tooling as increasingly important for both human developers and AI agents, while Cloudflare has committed $2 million in total toward open-source support, including a Vite ecosystem fund, and plans further improvements to Oxc, Rolldown, and an open-source self-hostable version of its Vite-native deployment platform.

Cloudflare is marking its 16th anniversary by launching The Cold Start, a live startup competition at Cloudflare Connect in San Francisco on October 19, 2026, inspired by its own public debut at TechCrunch Disrupt in 2010, which rapidly increased adoption and earned it recognition as the Most Innovative Company. Five eligible early-stage startups from the United States and Canada that have raised less than $10 million will be selected to deliver five-minute pitches and answer judges’ questions before Cloudflare co-founders Matthew Prince and Michelle Zatlyn and CTO Dane Knecht. The winner will receive $500,000 in Cloudflare credits, a San Francisco billboard, and an invitation to a VIP dinner, while all finalists will have travel covered. Cloudflare seeks clearly articulated, ambitious ideas rather than polished presentations, particularly products, infrastructure, interfaces, or software approaches made newly possible by changing technology. Applications, which ask companies to explain their product, funding stage, revenue, and use of Cloudflare, close October 2, 2026, and the initiative complements Cloudflare for Startups, which offers eligible companies up to $350,000 in credits to access global infrastructure, security, compute, and AI capabilities.

Cloudflare has released an experimental public preview enabling Rust applications built for the Emscripten wasm32-unknown-emscripten target to run on Workers through enhanced wasm-bindgen support, improving compatibility with native Rust libraries and Tokio-based applications. Developed through collaboration among Google, Cloudflare, Emscripten, wasm-bindgen, and Tokio contributors, the integration allows Emscripten and wasm-bindgen to coexist through a new configuration that supports Rust–JavaScript and C++–Rust interoperability. Cloudflare also added experimental approaches for integrating Tokio with JavaScript event loops, either through WebAssembly JavaScript Promise Integration or a proposed cooperative LocalEventLoop runtime that lets the host drive Tokio without blocking. More than 40 Emscripten contributions introduced a Node-compatible raw sockets layer supporting epoll, TCP, UDP, and Unix sockets, which Workers can provide through its Node.js compatibility APIs. As a demonstration, Cloudflare ran the Rust-based Pumpkin Minecraft server in a single-threaded Durable Object, adapting its threaded workloads into asynchronous tasks, persisting its files in Durable Object SQLite storage, and accepting multiplayer connections through Workers TCP ingress.

Cloudflare has expanded Kitesurf, its Cloudflare Workers-based browser designed for AI agents, with WebMCP support that lets agents call website functions directly instead of relying on fragile simulated clicks. The browser now supports additional web standards and APIs, including advanced CSS features, custom elements, JSON modules, import maps, URL-based module resolution, and improved iframe behavior, while its Web Platform Test coverage has grown to more than 730,000 passing subtests. Internal optimizations reduce JavaScript-to-DOM overhead, redundant timer and script work, memory use, and unnecessary font downloads, helping maintain or improve CPU and wall-clock performance despite broader compatibility. Kitesurf now has full integration with Cloudflare Browser Run through CDP, Playwright, Puppeteer, MCP, REST Quick Actions, and a Workers binding. Cloudflare also demonstrated a terminal-based version that renders pages through the Kitty graphics protocol or ANSI text and supports interactive scrolling and clicking. The company plans continued standards, performance, and rendering improvements, intends to open-source Kitesurf, and is offering it free in beta with account limits through its playground and Browser Run.

Cloudflare has introduced cf, an open-beta command-line interface designed to expand agent access from Wrangler’s roughly 280 commands to more than 3,000 Cloudflare API operations, responding to growing agent usage that reached 48% of Wrangler activity in the prior week. Built through Forge, a pipeline that generates commands from annotated OpenAPI schemas, cf standardizes command patterns across products and enables tasks ranging from Worker deployment and monitoring to domain purchases and security configuration. It uses JSON as its default output, provides natural-language command discovery through cf cli search, and offers validated forms for operations requiring user input. The tool also introduces cloudflare.config.ts, a TypeScript-based, type-checked configuration format intended to improve agent and editor assistance, support programmatic environment definitions, and eventually manage Cloudflare resources beyond Workers. Vite is the default development system, supplying its development server, plugin ecosystem, hot module replacement, and testing integration, while cf can still delegate to Wrangler for projects that rely on existing esbuild, Rust, or Python workflows. Cloudflare plans to maintain Wrangler for 18 months after cf’s beta concludes, while offering migration and initialization commands to help developers adopt the new CLI.

Cloudflare’s 16th anniversary message describes an Internet being rapidly reshaped by AI, with a surge in new websites since mid-2025 driven not primarily by low-quality content but by creators using AI-assisted development tools. The company says more than 7 million developers now build on its platform and predicts automated traffic, accelerated by AI agents and crawlers, surpassed human traffic in May 2026 and could become 1,000 times larger within five years. While agents can efficiently research products and services for users, Cloudflare warns that their extensive crawling imposes costs on websites and may favor established businesses with more available data, making it harder for new entrants and smaller local businesses to compete. The company advocates for an Internet where creators and businesses can be compensated when agents use their content, while introducing technology intended to reduce unnecessary crawler requests by serving updated material rather than repeatedly fetching unchanged pages.

Cloudflare has introduced Turnstile Spin, an agent-mediated workflow that helps users install, repair, or migrate to its free, privacy-focused CAPTCHA alternative, Turnstile. Standard Turnstile integration requires placing a client-side widget on a website and validating its token through the Siteverify API on the backend, but Spin enables coding agents such as Claude Code, Cursor, and Codex to identify relevant code, propose changes for approval, and implement both steps within the user’s own codebase. Spin supports new installations, recovery of widgets lacking server-side validation, and migrations from other CAPTCHA systems, while Cloudflare retains only the newly created Turnstile widget and does not receive application code. Available through the Cloudflare dashboard, Wrangler, or a public agent skill URL, the feature responds to growing demand for accessible bot protection amid wider AI-assisted application development. Since its July release, Spin has reportedly created more than 65,000 widgets successfully and generated over 30,000 copied prompts.

Cloudflare disclosed and remediated a vulnerability in its multi-tenant Containers and Sandboxes services after researcher Oren Yomtov of Accomplish reported it through HackerOne on September 4, 2026. The issue arose because Linux dm-thin storage pools had block zeroing disabled, allowing a newly allocated 64 KiB disk block to retain up to 60 KiB of data from a previously deleted container when only part of the block was overwritten. Researchers demonstrated that a Workers Paid customer could potentially recover residual filesystem metadata, database pages, directory structures, and application data from other customers placed on the same host, though they could not choose a target, host, workload, or specific data and could not access active disks or alter other customers’ data. Cloudflare removed the setting that bypassed block clearing, retired existing container disks, cleared cached image snapshots, and completed fleet-wide cleanup by September 19. Its review of historical disk-I/O telemetry found activity matching the technique only from the researchers and authorized Cloudflare validation, with no evidence of malicious exploitation or customer data compromise.

Cloudflare has launched Worker Previews, a feature that gives every Git branch its own production-like Workers environment with isolated code, configuration, URL, observability, and persistent state. Developers can create previews using `npx wrangler preview` or automatically through Git-connected Workers Builds, test changes on stable branch-specific URLs, inspect logs, metrics, errors, and traces, and safely use separate Durable Objects and Containers without affecting production or other branches. Previews begin from a reusable base configuration but allow branch-level overrides for resources such as databases, secrets, and API keys, and can use custom domains and Cloudflare Access to support realistic authentication, cookies, CORS, and OAuth testing. Cloudflare positions the feature as supporting an agent development lifecycle in which automated agents and developers can deploy, test, diagnose, revise, and verify changes before release, including browser-driven testing through tools such as Browser Run and Playwright. The company has used the capability internally for CloudflareOS and says future work will expand previews to multi-Worker applications, Queue consumers, Workflows, and long-lived staging or QA environments.

Cloudflare has introduced Vary support in Cache Rules across all plans to help sites cache multiple valid representations of a URL without excessive cache fragmentation. The HTTP Vary header tells caches which request headers may influence a response, such as Accept, Accept-Language, or Accept-Encoding, but raw header differences can create many separate cache entries even when an origin produces identical content. Cache Rules lets customers decide how Cloudflare handles each Vary field through normalization, which groups equivalent negotiation preferences; passthrough, which preserves exact header values for controlled variations; and bypass, which avoids caching for personalized or unpredictable fields such as cookies or user agents. Cloudflare recommends normalization for common content negotiation, while stressing that origins must consistently return correct Vary headers, including on error and fallback responses, to prevent incorrect cache reuse. The feature can be configured through the Cloudflare dashboard, Rulesets API, or Terraform, and is intended to provide a response-driven alternative to custom cache keys, which apply request dimensions to every covered response regardless of whether the origin varies on them.

Cloudflare has made Python Workers generally available, establishing Python as a fully supported language on its Developer Platform and allowing developers to deploy Python applications that integrate with services including Workers AI, R2, D1, Hyperdrive, Durable Objects, Queues, and Workflows. Built on Pyodide and WebAssembly, the platform now provides native, Pythonic access to Cloudflare bindings, eliminating prior JavaScript conversion requirements, while ASGI and WSGI connectors enable frameworks such as FastAPI, Django, Flask, and other compatible applications to run without managing servers. New socket support through the Workers connect API also enables PostgreSQL and MySQL drivers with Hyperdrive, as well as networking-dependent AI libraries such as OpenAI, LangChain, and MCP. Cloudflare has supported the broader Python-on-WebAssembly ecosystem by helping establish the PyEmscripten package platform through PEP 783 and extending build tooling, with the aim of increasing compatible native-extension packages. Available examples demonstrate uses including AI orchestration, image generation, real-time WebSocket stream processing, MCP servers, and retrieval-augmented generation systems, while future work will focus on performance, memory efficiency, and wider package support.

Cloudflare reduced the global memory use of its Pingora Backend Router by more than 100 TB by optimizing its consistent-hashing implementation, pingora-ketama, which routes cacheable requests to servers while maintaining stable cache placement. Consistent hashing uses virtual hash points to distribute workloads evenly and can weight servers by storage capacity, but Cloudflare’s feature-specific routing requirements created dozens of hash rings containing large numbers of points, consuming up to 6 GB of memory in some instances. Engineers first compacted each stored hash point from eight to six bytes by replacing a wasteful 32-bit server index with a 16-bit index and avoiding Rust structure-alignment padding, cutting hash-ring memory by 25%. Mathematical analysis and simulations then showed that the existing weighted configuration used far more hashes than necessary, with diminishing accuracy benefits and increasing collision risks from 32-bit hashes, allowing the team to reduce hashes per server by 90% without meaningful routing degradation. To prevent widespread cache invalidation and increased origin traffic, Cloudflare ran old and new rings simultaneously and migrated traffic gradually by data center while monitoring routing, errors, cache behavior, memory, and origin load. The resulting v2 ring, available in the pingora-ketama crate, also includes compact storage, faster sorting, configurable hash counts, and support for controlled parallel migrations.

Cloudflare describes how its Page Shield machine-learning system identified eight malicious JavaScript payloads across four live e-commerce operations that public scanners largely had not classified as harmful. The campaigns included mobile- and time-gated affiliate-commission hijacking through intercepted clicks, clickless affiliate attribution theft via hidden iframes, a repurposed Lnkr malware script that collected telemetry and could remotely execute arbitrary JavaScript, and a highly selective mobile-ad cloaker that disabled analytics and customer support while attempting to replace advertising and tracking identifiers. These scripts used obfuscation, typosquatted delivery domains, browser-state checks, IP and geographic filtering, delayed execution, cooldowns, and conditional remote loading to evade one-time crawls and analyst inspection while leaving storefronts apparently functional. The detection system analyzes JavaScript as a graph of code relationships using a graph neural network, validates suspicious findings with an LLM, and uses an ensemble of frontier models and human review for difficult cases, feeding results back into model training. The report argues that continuous client-side monitoring and behavioral analysis are necessary because malicious code may only activate for particular visitors, devices, campaign sources, locations, or browsing conditions.

Cloudflare has introduced resource-level access controls for Workers, allowing customers to grant users, agents, and API tokens permissions limited to a specific Worker rather than an entire account. Four roles—Metadata Read-Only, Content Read-Only, Editor, and Admin—provide graduated access for observing application data, reviewing code, deploying changes without deletion rights, or fully managing a Worker. Permissions can be applied across Developer Platform, product, or individual resource scopes, supporting least-privilege access for debugging, code review, CI/CD deployments, and administration. Routes and Custom Domains additionally require Workers Routes permission for the relevant zone, while Durable Objects inherit access from their associated Worker. Cloudflare’s APIs now provide documentation links in authorization errors to clarify needed permissions, and controls can be configured through the dashboard, API, or Terraform, including through user groups. Existing legacy Workers permissions remain supported without a planned deprecation date, though Cloudflare recommends migrating to the new roles and plans to extend the same model to products such as D1, R2, KV, and other Developer Platform resources.

Cloudflare has introduced a Disallow AI Training setting that allows website owners to remain visible in search indexes while signaling that their content should not be used for AI training, addressing the problem posed by mixed-use crawlers that perform both functions. The setting publishes a robots.txt preference, permits accountable mixed-use crawlers such as Applebot, Googlebot, and Bingbot to continue indexing sites for search, and blocks other training crawlers without affecting search visibility; selecting the broader Block option now stops mixed-use crawlers entirely, including their search activity. Cloudflare also replaces older AI bot controls with separate Search, Training, and Agent controls, migrates most existing settings automatically, and offers more restrictive recommended presets for advertising-supported sites. Apple, Google, and Microsoft have been designated Accountable based on existing or committed controls for AI-training opt-outs, AI-summary exclusions, transparency, and assurances that training choices will not harm search rankings, although Microsoft’s robots.txt support is targeted for early 2027. Looking ahead, Cloudflare plans to develop controls over the amount of website content used in AI summaries, noting that summaries may reduce traffic but can also send higher-intent visitors, making the appropriate choice dependent on each site’s business model.

Cloudflare has introduced automatic remediation policies for its Cloud Access Security Broker (CASB), enabling security teams to automatically respond to SaaS security findings by revoking risky file shares, sending webhooks, or forwarding events to SOC and SOAR tools. The feature builds on earlier manual remediation capabilities by allowing organizations to define event-driven response rules that execute as soon as findings such as publicly shared files, excessive OAuth permissions, or dormant credentials are detected. Policies can target selected SaaS vendors, integrations, and finding types, with direct remediation currently available for Microsoft 365 and Google Workspace file and folder findings, while webhooks can integrate with platforms including Slack, Teams, Jira, ServiceNow, Tines, and custom HTTP endpoints. Built using Cloudflare Queues, Workers, and Workflows, the system is designed for resilient execution, automatic retries, API rate-limit handling, and remediation completion within five minutes. Cloudflare One also provides audit and runtime logs to track policy changes, execution results, failures, and evidence of corrective actions for compliance, while planned Custom Findings support will let organizations tailor detection logic to their own requirements.

Cloudflare’s 1.1.1.1 resolver now validates DNSSEC signatures using ML-DSA-44, a NIST-standardized post-quantum algorithm, as an early step toward protecting DNS authentication from future quantum attacks and meeting Cloudflare’s stated goal of full post-quantum security by 2029. DNSSEC prevents forged DNS responses through a signed chain of trust from the root zone to individual domains, but current RSA and ECDSA signatures could be vulnerable to sufficiently capable quantum computers. ML-DSA-44 creates operational challenges because its 2,420-byte signatures and 1,312-byte public keys far exceed typical UDP DNS payload sizes, requiring reliable fallback to TCP or other transports and increasing bandwidth and processing demands. During the extended migration period, zones will need both conventional and post-quantum signatures for compatibility, creating a potential downgrade risk; 1.1.1.1 addresses this by requiring a valid ML-DSA-44 validation path when an authenticated parent DS record indicates post-quantum support. Complete protection will require adoption by authoritative DNS providers, registrars, registries, parent zones, and the DNS root, while Cloudflare plans further signing and DS-record support and will use large-scale measurements and test domains to assess deployment effects.

Cloudflare has rewritten workerd’s module registry, the open-source core of the Workers runtime, to improve Node.js compatibility, standards compliance, startup efficiency, and memory use, while retaining the existing registry for deployed Workers. Available through the explicit `new_module_registry` compatibility flag, the update supports `import.meta.url`, `import.meta.main`, and `import.meta.resolve()`, URL-based module specifiers with query strings and fragments, consistent `node:` built-in resolution, strict import-attribute validation, Node.js-style `require()` behavior for ES modules, uniform loading errors, lazy compilation, shared caching across V8 isolates, and WebAssembly source-phase imports. The changes address limitations of the prior filesystem-path-based registry, which eagerly compiled entire bundles and handled module features inconsistently, and may allow bundlers such as Rolldown to preserve more of an application’s module graph rather than transforming it extensively. Cloudflare also states that Workers now support all stable Node.js APIs relevant to serverless use, enable them by default, and allow application bundles up to 64 MiB on all plans without a compressed bundle-size limit.

Cloudflare has introduced Automatic Key Exchange, an extension of Automatic SSL/TLS that actively probes TLS 1.3 origin servers to identify their supported and preferred key-agreement algorithms, allowing Cloudflare to select an optimal algorithm in the initial handshake rather than relying on a universal X25519 default. The system prioritizes the post-quantum hybrid X25519MLKEM768 where supported, otherwise choosing compatible classical curves, and uses staged rollouts, monitoring, and daily rescans to avoid disruptions as origin configurations change. Cloudflare reports that the feature reduced HelloRetryRequest rates among scanned origins from roughly 52% to 3.7%, cutting more than 150 milliseconds from p90 handshake latency, while 99.2% of post-quantum-capable origin connections in the scanned cohort now complete in one round trip. Enabled by default for eligible domains, the feature can be controlled through the dashboard, including optional post-quantum-only or FIPS-compliant requirements, though enforcing unsupported requirements can prevent TLS connections. Cloudflare also provides Radar tools to test origin post-quantum readiness and plans future enhancements including per-origin preferences, on-demand scans, and automated post-quantum origin authentication to reduce certificate downgrade risks.

Cloudflare has announced invitation-only early access to Vulnerability Discovery and Remediation, a Managed Defense service designed to help customers identify, prioritize, and mitigate vulnerabilities in authorized codebases. Using OpenAI Daybreak models, including GPT-5.6 Cyber, the service maps production routes to relevant code, investigates potential weaknesses, validates findings, and proposes code patches and narrowly scoped WAF custom rules. Its prioritization incorporates source-code evidence with Cloudflare network context such as active routes, traffic volume, recent security events, and existing WAF protections, aiming to distinguish exposed, high-risk issues from less urgent findings. Customers retain control over implementation, as models cannot deploy patches or rules and every recommendation undergoes automated checks, Cloudflare team validation, and customer review. Investigations run through Cloudflare’s infrastructure and AI Gateway to OpenAI’s servers, with authorized data minimized, redacted according to engagement settings, and governed by access policies.

Cloudflare prototyped Cache Transcoding, a Pingora-based system that uses Zstandard compression to store eligible cache objects in compressed form, aiming to offset rising memory and disk costs while reducing inter-data-center bandwidth use. The system compresses unencoded, compressible text responses such as HTML, JSON, CSS, and JavaScript of at least 4 KiB when they enter the cache, retains the compressed representation on disk and across Tiered Cache transfers, and decodes it only before client delivery. Initial controlled tests found that eligible assets shrank by about 2.8 times on average, offering substantial effective cache-capacity gains in exchange for a modest CPU cost, with encoding paid once per cache fill and decoding required on each response. Already compressed media, precompressed content, binary data, range requests, and unknown-length responses are excluded because they offer limited savings or add complexity. Testing across more than one million requests and 10 cache servers validated cache-hit, cache-miss, and tiered-transfer behavior, though Cloudflare notes that broader real-world testing is needed before applying the measured compression ratio across its network.