August 2026 Summaries

Cloudflare has introduced Adaptive Intelligence, a bot detection engine designed to raise the cost of automated attacks rather than relying solely on static rules intended to block every attacker. The system continuously retrains machine-learning models on live network traffic and combines signals including TLS fingerprints, request patterns, session behavior, network reputation, browser telemetry, and challenge results to identify distributed, low-volume attacks that can evade conventional rate limits. Planned capabilities include automatically generating short-lived “disposable” detections that are difficult for attackers to reverse-engineer and retaining knowledge of past attack patterns after individual rules are retired. Updates are tested in shadow mode and rolled out gradually to limit false positives affecting legitimate users, with performance monitored through measures such as challenge outcomes, precision, recall, and customer feedback. Adaptive Intelligence is intended to work alongside Cloudflare’s Precursor behavioral-validation system, and Enterprise Bot Management customers can receive it by enabling Auto Update Machine Learning in the dashboard.

Cloudflare has launched BotBase for Operators, expanding its bot directory service to give bot operators greater visibility and control over submissions while helping website owners manage automated traffic. Available through the Cloudflare dashboard, the new interface includes a searchable bots directory, a submission form, and submission history showing whether entries are awaiting review, accepted, or rejected, with explanations for decisions and classification changes. Operators can now edit existing submissions or cancel pending ones, helping keep verification information such as IP lists, reverse DNS, or Web Bot Auth details current. The revised form uses a taxonomy that captures a bot’s functions, its intended use of accessed content, and whether it is operated directly or as an intermediary platform, aligning bot declarations with website Content Signals preferences. Cloudflare also automated many review checks to address a sevenfold rise in submissions since 2023, aiming to speed approval while routing complex cases for manual review. Future plans include bot ownership tools, visibility into how sites treat bots, and mechanisms for operators and websites to communicate about access and value.

Cloudflare’s Big Pineapple DNS platform, which supports services including 1.1.1.1 and DNS Firewall, holds more than 250 billion cache entries and reduced its cache memory footprint through five Rust-focused storage optimizations. The changes replaced mutable vectors and strings with fixed-size boxed storage, consolidated DNS response sections, omitted record owner names when they match the queried domain, avoided oversized enum representations, and stored record data as compact contiguous wire-format bytes. Benchmarks using representative DNS traffic showed that these modifications cut per-entry memory use from 953 bytes to 420 bytes and allocations from 1.1 KB to 461 bytes, while improving insert throughput by 43% and reducing lookup latency by 19%. During a May-to-July 2026 production rollout, p99 process memory usage declined from 9.3 GB to 5.3 GB per instance, contributing to an estimated 100 TB reduction in fleet-wide working memory; Cloudflare plans to use the reclaimed capacity to expand caches and improve hit rates.

Cloudflare migrated and redesigned its blog on August 12, moving from its previous CMS to EmDash, an Astro- and Cloudflare-focused content management system, as part of its “Customer Zero” practice of using and improving its own products internally. The project identified and addressed early-platform gaps involving scheduled publishing, large-scale media and content search, localization, SEO, security policies, and editor usability, while performance testing informed an architecture based on Cloudflare Workers, Workers Cache, Workers KV object caching, and Hyperdrive integration with PlanetScale. The new design aligns with Cloudflare’s Kumo system, adds accessible light and dark modes, relocates email subscriptions to post endings, and introduces page navigation and discussion-sharing sidebars. A proxy Worker enabled a gradual, reversible rollout from 1% to 100% of traffic with fallback for server errors. Cloudflare reports that the new cached system delivers more consistent latency, serves most static assets and many requests from cache, handled up to 850 requests per second in early production, supported nearly 3 million pageviews during Agents Week, and absorbed a 28,000 RPS DDoS attack. The migration also added Model Context Protocol servers that let AI agents search blog posts and enable authors to manage EmDash content through agent-based tools.

Cloudflare is introducing Bot Preference Sync, a feature available across all plans that automatically updates a website’s robots.txt file to match AI bot policies configured in its dashboard for Search, Agent, and Training traffic. The tool is intended to reduce conflicts between stated crawler preferences and edge-enforced bot controls, while allowing site owners to choose policies that fit their business models, such as maximizing AI-driven discovery or preventing model training on ad-supported content. Training can be disallowed without necessarily affecting search indexing for verified mixed-use crawlers that meet Cloudflare’s transparency requirements, including honoring no-training directives, providing opt-outs from AI summaries, and offering page-level usage visibility. Bot Preference Sync prepends Cloudflare-managed directives to existing robots.txt content and updates bot lists using Cloudflare’s verified bot data, though customers with specialized exceptions or custom rules can disable it and manage their files independently. New publishing customers can select an ad-monetization option that defaults to disallowing training, while other new customers receive no default blocks, reflecting Cloudflare’s broader emphasis on transparency, consent, discoverability, and site-owner control over AI crawler access.

Cloudflare has introduced OAuth scope customization, allowing developers to designate certain requested permissions as optional so users can deselect them during authorization rather than accepting an all-or-nothing access request. The feature addresses increasingly granular permission needs for SaaS integrations, internal tools, CLIs, and AI agents such as MCP servers, while preserving existing behavior for clients that do not enable optional scopes. Required and optional status is evaluated only among scopes requested in each individual authorization flow, keeping consent screens focused on the access needed for a particular task. When users decline optional permissions, issued access tokens include only the scopes granted, requiring developers to design applications that can function safely with partial authorization. Cloudflare also plans to expand account- and zone-level roles, API token roles, membership options, and OAuth scopes across nearly all of its products, further supporting fine-grained access control.

Cloudflare reassessed remote Spectre risks in its Workers platform after advances in attack-stabilization techniques and discovered that a limitation in its Dynamic Process Isolation (DyPrIs) defense enabled a controlled end-to-end cross-tenant memory-leak demonstration in production. Researchers overcame frozen local timers, shared-resource noise, cache-eviction challenges, and execution limits by combining speculative type-confusion gadgets, cache-state amplification using the L1 cache’s PLRU policy, remote WebSocket-based timing, co-location through Worker subrequests, and persistent Durable Objects. The proof of concept leaked a deliberately planted JWT token from a co-located Worker at up to 12 bits per second with more than 99% accuracy, substantially faster than an earlier attack, although Cloudflare found no evidence of exploitation. The attack avoided DyPrIs detection because isolation occurred only after invocation completion and I/O-heavy WebSocket activity distorted its hardware-counter detection metrics. Cloudflare reports that the issue has already been mitigated through improvements to DyPrIs, deployment of the V8 memory sandbox, and hardware-assisted in-process isolation using Memory Protection Keys, which creates stronger boundaries between isolate heaps while not fully eliminating the broader class of Spectre-related risks.

Route leaks occur when BGP announcements travel beyond their intended customer-provider or peer-to-peer relationships, potentially causing traffic misdirection, congestion, latency, and outages. RFC 9234 addresses this problem by introducing BGP Roles, which let neighboring networks declare and validate their relationship, and the Only to Customer (OTC) attribute, which marks routes that should only continue toward customers and allows routers to reject leaked paths automatically. Cloudflare assessed adoption through public routing data and direct observations from its global peering network, identifying relatively limited deployment, with 67 peer ASes observed sending OTC attributes and route servers appearing among the earliest adopters. Its experiments also found that several networks stripped OTC attributes despite their required transitive handling, including Tier-1 providers GTT and Arelion; these providers said the practice originated as a defense against historical BGP error-handling risks, and Arelion subsequently began preserving OTC following Cloudflare’s outreach. Because a small number of high-tier networks can affect a large share of Internet paths, retaining OTC is important for effective route-leak prevention during partial deployment. Cloudflare recommends that operators configure BGP Roles where supported, coordinate deployments during maintenance windows because sessions must reset, and encourage vendors without RFC 9234 support to implement it.

Cloudflare announces Cloudflare One capabilities for identifying, monitoring, and controlling Model Context Protocol (MCP) traffic, addressing the heightened risk of AI agents that can make nondeterministic decisions and execute actions at machine speed using existing organizational permissions. MCP tool calls can be governed at the client, network, and server layers, each offering different visibility and enforcement capabilities, while network inspection provides broad coverage of remote traffic but cannot observe local, off-network, or undecrypted connections. Gateway can now classify TLS-inspected MCP requests using protocol headers, expose related users, hosts, and traffic sources in logs and a dedicated dashboard, and enable policies that allow or block detected traffic, including direct connections that bypass approved Cloudflare MCP Portals. MCP Portals provide managed endpoints with identity controls, curated tool catalogs, logging, data-loss-prevention routing, and support for pre-registered OAuth clients, helping organizations distinguish and address both unapproved “shadow MCP” servers and direct bypasses of approved services. The update also notes planned private-network connectivity for Portal-managed MCP servers and adds support for the newer stateless MCP protocol in Cloudflare’s Agents SDK while retaining compatibility with legacy implementations.

Cloudflare has introduced Access for Workers, allowing organizations to enforce authentication directly on individual Workers or across an entire account so applications, previews, and deployments remain private by default. Unlike earlier hostname-based policies, Worker-level protection automatically applies to all associated custom domains, routes, workers.dev subdomains, and preview URLs, while organizations can choose to secure previews, production traffic, or both and exempt specific public Workers when needed. Access integrates with existing identity providers, email or group restrictions, and service tokens, and provides authenticated user details such as email, name, and groups to Worker code through ctx.access.getIdentity(), removing the need for developers to validate JWTs manually. Local development can simulate authenticated identities through Wrangler configuration, while Workers for Platforms can protect all tenant deployments by applying Access to a dispatch Worker. Cloudflare says the feature was enabled by its Rust-based FL2 proxy architecture, which allowed routing to occur before Access authentication, and it is now available to all users.

Cloudflare has made its Certificate Transparency Monitoring service generally available after redesigning it to suppress alerts for certificates that Cloudflare itself issues and renews, reducing routine notification noise across more than 650,000 monitored domains. The change addresses frequent alerts caused by short-lived Universal SSL, Advanced Certificate Manager, Total TLS, and backup certificates, whose issuance must be publicly recorded in Certificate Transparency logs. Cloudflare connected its certificate-ordering and alerting systems using an SHA-256 hash of the certificate’s SubjectPublicKeyInfo public key, which is generated early, remains consistent across certificate lifecycle stages, and can be independently recalculated from log entries. This enables the service to recognize and suppress Cloudflare-managed pre-certificates and final certificates while continuing to alert customers about externally issued or uploaded custom certificates. Updated alert emails now identify the affected hostname, provide certificate details, and link to the Cloudflare dashboard, while planned integration with Cloudflare Notifications may add routing to webhooks, PagerDuty, and additional email destinations.

Cloudflare Radar data showed that the August 12 solar eclipse across Europe coincided with sharp, precisely timed declines in HTTP traffic as people paused online activity to watch the event. Comparing five-minute traffic intervals with the median traffic levels from the prior three Wednesdays, the analysis found that decreases generally peaked around maximum solar obscuration and were largest in locations along or near the path of totality, including Iceland, Spain, and Portugal, where declines ranged as high as roughly 15% to 30% or more in some measurements. Countries experiencing shallower eclipses, such as Sweden, Denmark, Poland, and Switzerland, saw limited changes, while Norway and Sweden recorded slight increases above baseline at times. The findings account for geographic variation, local conditions, and normal daily traffic patterns, but the close alignment between eclipse timing, obscuration depth, and traffic troughs indicates that the shared celestial event was the main cause of the temporary reduction in Internet activity.

Cloudflare’s H1 2026 DDoS Threat Report, based on telemetry from its network, describes a substantial rise in network-layer attacks, including 935 attacks exceeding 1 Tbps and a more than sixfold quarterly increase in such hyper-volumetric events during Q2. It reports mitigating 23.2 million network-layer attacks and 29.64 trillion HTTP DDoS requests from January through June, while noting that most attacks remained below 500 Mbps and lasted under 10 minutes, making automated protection important because even brief attacks can cause extended disruption. DNS-related attacks represented 34.3% of network-layer activity, with DNS floods growing sharply and CLDAP amplification attacks increasing 580% quarter over quarter. Geopolitical events appeared to coincide with targeting patterns: media organizations were the most attacked industry, government-sector targeting rose following Operation Epic Fury, and Turkey became the third most-attacked location amid the Ankara NATO Summit buildup. China and the United States were the most targeted countries, while Brazil became the leading apparent source country for DDoS traffic. The report also links a decline after April’s activity peak to the multinational Operation PowerOFF crackdown on DDoS-for-hire services and highlights Cloudflare’s global capacity, automated mitigation systems, and botnet threat-sharing program.

Cloudflare’s Agents Week outlined its vision for an “Agentic Internet,” in which AI agents operate alongside people using infrastructure designed for autonomous software. Across five days, the company announced and discussed agent-focused runtimes, cross-language Workers communication, model-serving improvements, usage tracking, TCP and gRPC support, and an Agent Development Lifecycle featuring tracing, replay, approvals, programmable wallets, and AI-assisted CI/CD. It also presented security and governance concepts, including an Agent Access Model, identity-aware analytics, WriteGuard controls for MCP servers, and the open-source Cloudflare OS platform. Later sessions focused on standards and tools for making websites discoverable, callable, and payable by agents, including WebMCP, MCPv2, AI Search, AEO practices, and Kitesurf, an agent-oriented browser. The week concluded with updates to Cloudflare’s AI control plane, bot-trust analysis, community and open-source funding programs, and Radar Researcher, positioning these offerings as components of a broader platform for building, securing, observing, and operating AI-native applications.

Cloudflare announced that its Cloudflare for Government offering has received FedRAMP High authorization, sponsored by the National Institute of Standards and Technology, enabling it to serve U.S. agencies handling highly sensitive unclassified information related to areas such as law enforcement, emergency services, financial systems, and national security. FedRAMP provides standardized federal security assessment, authorization, and continuous-monitoring requirements, and the High level represents a substantial step beyond Cloudflare’s 2022 Moderate authorization because failures could have catastrophic consequences. Rather than operating a separate government-only cloud, Cloudflare says its government service uses its existing global network and software stack while relying on its Data Localization Suite to ensure FedRAMP High traffic processing and storage remain within U.S. data centers. The company says this approach gives government customers access to current Zero Trust, DDoS protection, performance, and developer capabilities without waiting for isolated government-platform releases. Cloudflare also plans to use the systems developed for FedRAMP High as the basis for pursuing Department of Defense Impact Level 4 authorization for controlled unclassified data.

Cloudflare outlines a web security strategy that evaluates automated and human traffic through continuous behavioral analysis rather than treating bots as inherently harmful or relying on one-time checks such as CAPTCHAs. It distinguishes short-term risk from trust built through reputation and transparency, describing BotBase as a directory that tracks verified and unverified bots and agents based on declared identity and observed behavior. Its Precursor system uses CDN-injected client-side JavaScript to assess behavior throughout a session, including cursor movement and transitions between human and agentic activity; Cloudflare reports 206 million evaluation events across 73,438 zones during a 24-hour period. The company also previews Adaptive Intelligence, a self-adjusting bot-detection engine intended to update continuously as traffic patterns evolve, and proposes mitigations such as randomized responses, AI Labyrinth traps for unauthorized crawlers, and queues for legitimate automated agents. These capabilities are intended to give site owners more granular control over automated traffic while allowing beneficial bots and user-directed agents to operate.

Cloudflare has introduced an expanded community program to recognize and support people who help others build with its platform, contribute to open-source software, and strengthen developer communities. The program includes Cloudflare Ambassadors, who may organize events, lead student groups, create educational content, and receive resources such as credits, marketing materials, technical support, and recognition in Cloudflare’s online spaces; applications for the 2026–27 cohort are open through September 6, with selections announced by October 5. A second track, Cloudflare Community Engineers, will provide grants from an additional $1 million fund over two years for contributors to open-source projects related to Cloudflare’s ecosystem, initially including projects such as Astro, Agents SDK, EmDash, Hono, and Vinext, with applications opening later. Cloudflare is also forming a Discord committee involving ambassadors and staff to connect users with experts, facilitate community discussions, and improve the server’s usability, while using automation to reduce spam and malicious content.

Cloudflare has beta-launched Radar Researcher, an AI-powered assistant that lets users explore Cloudflare Radar’s global Internet data through plain-language questions, returning explanations, interactive charts, suggested follow-ups, and auditable traces of the datasets and API calls used. Available across Radar pages, it can analyze existing charts using their screenshots, raw API data, and active filters, helping users investigate topics such as national Internet quality or disruption timelines without manually navigating dashboards or APIs. Built entirely on Cloudflare’s developer platform, the tool uses Workers, the Agents SDK, Durable Objects, Workers AI, AI Gateway, and a unified MCP server that enables it to discover and query Radar’s APIs dynamically rather than relying on hard-coded endpoints. Its chart system keeps raw data separate from model-generated prose, allowing the frontend to render precise, native Radar visualizations from API-backed references. Cloudflare has also added WebMCP support to Radar so browser-based agents can directly use site functions such as filtering, searching, URL scanning, and domain reporting, while maintaining normal functionality for browsers without agent support.

Cloudflare is converging AI Gateway and Workers AI into a unified platform that lets developers access Cloudflare-hosted and external AI models through shared Workers bindings and REST API endpoints, with centralized observability, security, logging, and billing. Users can route Workers AI requests through an automatically created default gateway to receive request and response logging, token tracking, latency and error metrics, and cost attribution without prior dashboard configuration, while named gateways remain available for customized routing and caching. AI Gateway credits can now pay for Workers AI as well as third-party providers, with unified prepaid billing and elevated Workers AI rate limits for users of this path. Cloudflare also plans model-first routing, which would select among vetted providers hosting a requested model to provide transparent load balancing and failover while honoring requirements such as zero data retention, followed by smart routing that classifies tasks and chooses suitable models automatically for performance and cost efficiency.

The MCP 2026-07-28 specification redesigns the Model Context Protocol as a fully stateless HTTP-based protocol, removing required session handshakes, session IDs, and persistent streams that previously complicated deployment, scaling, and recovery for remote MCP servers. Requests now carry their own version, client identity, and capabilities, while Multi Round-Trip Requests replace stream-dependent elicitation by allowing servers to request additional user input and clients to retry operations with it. New HTTP headers expose MCP methods and tool names to gateways, rate limiters, and security tools, and caching hints improve reuse of tool, prompt, and resource catalogs. The release also strengthens authorization through preferred pre-registered clients, Client ID Metadata Documents, issuer validation, and stricter token audience requirements, while deprecating Dynamic Client Registration and several legacy capabilities under a formal lifecycle policy. Cloudflare says its updated SDKs and Workers-based createMcpHandler enable MCP servers to run as ordinary stateless workloads without Durable Objects unless an application itself requires coordinated state, with compatibility paths for older clients and production adoption by organizations including Sentry and Linear.

Cloudflare argues that businesses increasingly need to optimize for AI agents and assistants that discover, evaluate, and recommend products or services before customers visit a website. Its Agent Readiness dashboard feature evaluates whether a site is accessible and understandable to agents by checking elements such as robots.txt, sitemaps, machine-readable Markdown, metadata, APIs, authentication, and emerging agent protocols, then provides evidence, improvement guidance, and Cloudflare setup links. The new Answer Engine Optimization tool measures how often leading AI assistants such as Claude and GPT mention or cite a business in relevant category queries, reporting metrics including citation rate, prominence, mention rate, share of voice, and industry fit relative to competitors. Cloudflare builds category-level benchmarks from repeated model queries and combines them with site-specific operator activity data, including crawling, referrals, and errors, to help owners assess agent traffic and test improvements. The company positions these tools as a way for businesses to treat AI agents as an important new audience and increase their likelihood of being found, understood, trusted, and recommended.

Cloudflare has launched a developer preview of WebMCP, a feature that lets sites expose browser-based tools to AI agents through the emerging WebMCP standard in Chrome without requiring code changes or origin redeployment. Enabled through the Cloudflare Dashboard, the service injects a same-origin bridge script at the edge that registers selected tool packs with a browser’s document.modelContext interface, allowing agents to interact with site capabilities directly rather than navigating pages designed for humans or relying on scraping. The preview includes a Content Credentials pack that locally examines C2PA image provenance metadata and a Site MCP Server pack that exposes tools from a site’s existing MCP endpoint while preserving the visitor’s same-origin session. All current tools run in the visitor’s browser, though Cloudflare plans future packs that can use edge services such as Workers AI and AI Search. Site owners can test the feature with Cloudflare BrowserRun and are invited to provide feedback as the experimental offering develops.

Cloudflare has introduced developer experience improvements to AI Search that automate the assembly of services such as Workers AI, Vectorize, R2, Browser Run, and AI Gateway into a managed search solution for agents and applications. Users can index structured and unstructured files or owned websites, crawl sites without sitemaps through a Discover option, combine multiple search instances under a namespace, and expose unified public /search and /mcp endpoints that can use custom domains or Cloudflare Access for private access. Cloudflare demonstrates the platform through its Dev Stack MCP server, which searches current, cited content across ten Cloudflare-owned documentation and community sources, as well as through search on Cloudflare.com, Developer Docs, and its EmDash-based blog and CMS integration. AI Search supports hybrid semantic and keyword retrieval, respects robots.txt and bot controls through its Cloudflare-AI-Search crawler identity, and remains free during beta, while preview pricing proposes charges for ingestion, storage, and queries with free monthly allowances and no embedding or reranking fees when customers use selected default Workers AI models.

Cloudflare argues that AI agents are becoming a major new type of web visitor, creating billions of redundant requests while also offering potential commercial value because they act on behalf of paying users. It proposes an “Agentic Internet” built around four capabilities: making content readable through agent-friendly formats such as Markdown for Agents and the lightweight Kitesurf browser, discoverable through AI Search and Agent Engine Optimization, callable through WebMCP and code-based tool access, and payable through wallets, x402, and its Monetization Gateway. The company emphasizes open standards including Web Bot Auth, PACT, MCP, and x402 to let sites authenticate legitimate agents, control access, enable anonymous trust signals, expose actions directly, and charge for content or APIs. Cloudflare presents these tools as a way for publishers, merchants, developers, and nonprofit providers to distinguish productive, compensating agents from extractive scrapers, while avoiding an Internet controlled by a small number of proprietary platforms.

Cloudflare has introduced Kitesurf, a lightweight, agent-focused browser built entirely on Workers and offered free in beta through Browser Run, aiming to reduce the cost and resource demands of Chromium-based automation. Designed for AI agents rather than human browsing, it prioritizes machine-readable content, scalability, isolation, token efficiency, and safety against risks such as prompt injection over features like tabs, extensions, pixel-perfect rendering, or persistent sessions. Kitesurf uses Rust and WebAssembly components alongside Cloudflare Workers features including Dynamic Workers, Durable Objects, RPC, and isolated outbound networking, with an architecture centered on a session-managing Engine, per-page PageScript isolates, and stateless PageRenderer workers. Development relied heavily on AI-assisted coding guided by Web Platform Tests, real-site integration tests, and visual regression comparisons with Chromium; the browser currently passes more than 215,000 web-platform tests. In reported benchmarks, Kitesurf used roughly three to seven times less CPU and memory than Chromium for screenshots and HTML extraction, although it was about 1.7 times slower in wall-clock time. It supports CDP-compatible tools such as Puppeteer and Playwright and is suited to compatible one-shot tasks including page extraction, screenshots, and PDF creation, while Chromium remains preferable for video, WebGL, sophisticated bot challenges, and long-lived authenticated sessions.

Cloudflare CIO Sam Rhea describes how rapidly improving AI agents prompted the company to develop Cloudflare OS, an internal platform for enabling employees to build and deploy AI-powered workflows while protecting internal and customer data. The effort is guided by principles including focusing on customer-serving problems, making AI accessible beyond engineers, holding humans accountable for agent outputs, supplying organizational context, and preserving existing permission boundaries. For engineers, Cloudflare created an Engineering Codex and AI review agents that have flagged hundreds of thousands of potential issues and blocked thousands of code merges, while nontechnical teams initially used a “magic AI email” service to identify recurring automation needs. Cloudflare OS evolved from a browser-based workspace that runs reusable skills with controlled access to company systems into a platform where users can describe workflows in natural language, generate secure applications and agents, and schedule or trigger them without managing API keys or deployment pipelines. The system uses Zero Trust authentication, MCP-based connections, custom permission controls, AI Gateway logging and data-loss protections, and isolated environments to govern access and model usage. Cloudflare reports that thousands of employees now use the platform weekly, with sales teams estimated to have saved more than 10,000 hours in one month and users creating more than 4,000 targeted apps and tools.

Cloudflare announces that Gartner recognized it as a Visionary in both the 2026 Magic Quadrant reports for SASE Platforms and Security Service Edge, positioning the recognition as support for its unified Cloudflare One architecture. The company argues that enterprises increasingly need SASE platforms able to address fragmented security tools, unmanaged AI agents, shadow applications, post-quantum cryptographic risks, and complex pricing models. It says its global network and single-platform design provide integrated zero trust controls, AI-agent governance, AI cost controls, post-quantum encryption, data protection, and programmable security capabilities through Cloudflare Workers. Cloudflare predicts that SASE will evolve into an agile governance layer that secures AI-built applications and autonomous agents, supports post-quantum migration, and consolidates security and connectivity on unified infrastructure, while noting its target of a fully quantum-secure SASE platform by 2028. Gartner’s reports represent the opinions of its analysts and do not constitute endorsements or factual guarantees.

Cloudflare has announced an open beta integration between AI Gateway and Cloudflare Access, along with the general availability of User Insights for AI Gateway customers, to improve visibility, identity-based governance, and cost management for organizational AI use. AI Gateway centralizes requests to AI providers and developer tools, while the Access integration authenticates users through SAML identity providers, attaches verified user IDs to requests, and enables per-user spending limits and access policies. User Insights establishes behavioral baselines for people and agents by comparing session costs against each account’s rolling 30-day p95 usage and an organization-wide p99 cost threshold, highlighting unusually expensive sessions while filtering out routine high usage or insignificant deviations. The system is intended to help administrators investigate possible compromised credentials, runaway agents, or inefficient AI practices without automatically blocking users. Cloudflare also plans task-based model routing to reduce costs and prompt classification to provide more context about how AI is being used, including potential insider-risk signals.

Cloudflare introduced WriteGuard, a shared policy, attribution, and auditing layer designed to manage AI agent write actions across Model Context Protocol (MCP) servers after recognizing that agents operating under employee credentials can make rapid, difficult-to-distinguish changes. MCP servers connect AI clients to external tools and systems, and Cloudflare’s internal portal now links dozens of initially read-only servers for services such as Jira, GitLab, wikis, and operational tools. WriteGuard assigns each tool a risk tier, can allow, enrich, or block calls before execution, adds agent-session labels to supported downstream writes, and records scrubbed asynchronous audit events containing details such as the user, tool, outcome, client, and duration. In a GitLab example, it permits read-only merge-request queries, labels and logs agent-created comments, and blocks merge actions classified as critical because they may trigger deployments. Cloudflare retained employee-based permissions rather than creating separate agent accounts, using WriteGuard to preserve both human accountability and agent-specific context. Having deployed the system internally, the company is launching a limited private beta for MCP server portals to test risk classifications, attribution formats, and audit requirements before broader availability.

Cloudflare has open sourced Cloudflare OS, an enterprise platform intended to give employees browser-based AI agent workspaces grounded in company-specific context, skills, systems, and policies. Developed from an internal deployment used across Cloudflare, it enables users beyond engineering to research information, create connected documents and presentations, automate repeatable workflows, and build collaborative full-stack applications. Its security model begins with no default access and uses Cloudflare Access, capability-based resource bindings, and service-specific Gatekeepers to keep credentials isolated, limit actions and data exposure, log observations, and enforce permissions when workspaces, outputs, or apps are shared. Apps run as isolated Dynamic Workers with durable state and can be shared either for real-time collaboration or as reusable blueprints without transferring data or credentials. The platform supports multiple AI models through Cloudflare AI Gateway, allowing organizations to control model selection, costs, budgets, and usage attribution. Organizations can deploy and customize the open-source core in their own Cloudflare accounts, integrating internal systems, governance policies, interfaces, and organizational knowledge, while Cloudflare plans managed dashboard availability and additional integrations.

The Agent Access Model (AAM) proposes extending Zero Trust principles to AI agents, whose short-lived, machine-speed tasks and multi-system access needs make traditional human- and service-account controls inadequate. Rather than trusting a task after initial authorization, AAM evaluates every action against the agent’s identity, the authorized task, and the task’s accumulated security state, with capabilities only able to narrow over time. Its reference architecture combines short-lived, task-scoped, sender-constrained credentials; a task-scoped access engine; enforcement in both the agent harness and network egress layer; a Trust Ratchet that removes permissions after protected events such as sensitive-data reads; and logging and review systems for refining future permissions. The model emphasizes that prompts cannot enforce security boundaries, since agents may be manipulated by injected content, and that enforcement must occur where tool calls and network requests are actually executed. A finance reconciliation example illustrates how a protected report can trigger restrictions that block later attempts to send data to vendor support while allowing a narrowly defined internal summary. Activity records collected from external enforcement points support auditing, incident investigation, and evidence-based changes to task templates without expanding permissions during an active run. Human approvals are intended for exceptional, high-risk decisions rather than routine actions, reducing approval fatigue. The paper identifies shared or multi-user agents as an unresolved problem because data, context, cached outputs, and generated responses may carry differing permissions across users, requiring stronger end-to-end provenance and authorization mechanisms.

Cloudflare is enhancing its developer platform to streamline the process of storing, building, testing, and deploying code directly on its infrastructure. With the introduction of Artifacts for versioned code storage and a CI SDK built on Cloudflare Workflows, developers can now run continuous integration (CI) pipelines seamlessly on Cloudflare. This system automates various stages, such as code compilation, linting, type-checking, unit testing, and conditional deployments, all within isolated environments. A notable feature is the inclusion of a self-healing AI agent that can detect and fix errors during builds, thereby reducing manual intervention. Developers can customize their CI/CD pipelines using TypeScript for greater flexibility compared to traditional YAML configurations. The platform supports both platform-managed and custom CI, allowing concurrent operations within the same namespace, simplifying the management of CI/CD processes across different applications and customer needs. Cloudflare's focus on Workflows not only enhances observability and retry mechanisms but also integrates with broader infrastructural components like R2 for caching and durable execution, all while promising future features such as direct integrations, gradual deployments, and enhanced monorepo management.

Cloudflare has introduced a new feature called Cloudflare Agents, designed to facilitate the deployment and management of hosted agents by providing enhanced observability. This initiative leverages Cloudflare's nine years of experience in building a developer platform, offering tools such as model access, durable runtime, orchestration, sandboxed execution, and persistent storage. A key component is the launch of agent tracing, which allows for direct insight into agent behavior, measuring every model call, tool execution, and token usage. This tracing is compatible with OpenTelemetry-compatible agent harnesses like Think, Flue, and AI SDK, enabling users to visualize and analyze agent operations to improve performance. The Cloudflare dashboard now features a dedicated Agents view to monitor and debug agents, offering functionalities like session replay and trace inspection to understand and optimize agent decisions. The integration supports exporting telemetry data to any OTLP-compatible provider, providing a feedback loop for improving agent quality, performance, and cost. While currently in beta, tracing will eventually be integrated into existing Workers Observability pricing by 2026.

Cloudflare has introduced automatic OpenTelemetry trace capturing for local Worker invocations using wrangler dev and vite dev, allowing developers to debug more efficiently without additional setup. This feature leverages the Local Explorer API, a local debugging tool that provides agents with access to traces and logs, enabling them to identify and resolve issues in the development environment before deployment. Developers can use the Local Explorer's browser-based interface or REST API to inspect local resources, trace data, and binding states, ensuring a streamlined debugging process. The integration with Cloudflare Workers and Miniflare offers a seamless experience by capturing runtime events and assembling them into OpenTelemetry traces, which can be queried and analyzed locally. This enhancement underscores Cloudflare's investment in improving local development workflows, enabling developers to reproduce errors, inspect telemetry data, and verify fixes efficiently without deploying code or adding temporary logs.

Engineering managers have historically focused on optimizing collaboration among programmers within the Software Development Lifecycle (SDLC), but the advent of AI has shifted the slowest phase, implementation, to become the fastest, creating challenges for subsequent stages. Cloudflare proposes a transformative approach with the Agent Development Lifecycle (ADLC), aiming to revolutionize software production by empowering agents to manage entire processes autonomously. This transition involves reimagining software development as a factory model, where agents autonomously handle tasks from bug reports to feature implementation. Cloudflare's innovations, such as @cloudflare/ci and Cloudflare Workflows, provide the infrastructure for this, enabling agents to manage complex workflows, ensure observability via OpenTelemetry, and maintain high standards of software production. The goal is to alleviate human bottlenecks in software production, allowing developers to focus on creative and strategic aspects while agents handle technical execution, thus paving the way for a new era of automated software development.

Cloudflare is addressing the challenges AI agents face in interacting with APIs by introducing Cloudflare Wallets, which offer a stable identity and payment method for AI agents to explore APIs and content more effectively. These wallets allow agents to make micropayments using the x402 protocol, providing a frictionless way to try out various services without traditional account setup hurdles. Cloudflare Wallets come in two forms: Account Wallets for human users to manage funds and delegate spending to Virtual Wallets used by agents, and Virtual Wallets that operate via API keys with spending limits to ensure controlled exploration. This initiative is part of Cloudflare's broader effort to foster agentic commerce by enabling automated transactions between AI agents and merchants, while proposing a human-readable identifier system to improve agent identity management. The company envisions a future where these tools will create a seamless, headless marketplace on the Internet, allowing agents to perform transactions autonomously with minimal human intervention.

Cloudflare has developed the Cloudflare Codex, a comprehensive set of engineering standards designed to streamline the development process by providing consistent guidance to engineers and AI agents. This system replaces the previous fragmented approach where guidance was dispersed in various forms and locations, making it inefficient for engineers to locate and verify. The Codex is organized into domains, each maintained by domain owners, and follows a Request for Comments (RFC) format to ensure standards are clear and authoritative. AI agents utilize the Codex to review code, technical designs, and incident reports, flagging violations and ensuring compliance. This approach has significantly improved the consistency and speed of engineering processes, with the AI code reviewer alone flagging nearly 230,000 violations and blocking 16,000 merges. The Codex has also been integrated into various tools, allowing for quick linter checks and local code reviews via a command-line interface. Cloudflare plans to extend the Codex model to other teams, such as product and security, and aims to enhance the agents' capabilities to propose fixes autonomously while maintaining human oversight.

Amidst discussions about the feasibility of AI-driven software factories and concerns over open-source maintainer burnout, a practical solution has emerged through the development of an automated triage pipeline on the Astro repository. This system, developed into the open framework Flue, automates the process of reading, reproducing, diagnosing, and verifying bug reports, significantly reducing the number of open issues from over 200 to about 30, with the expectation of reaching zero soon. The approach involves using isolated AI subagents within GitHub Actions to manage issue triage without resorting to dismissive practices like auto-closing tickets. The success of this method has led to more meaningful community interactions and has highlighted areas for improvement in code documentation and test coverage. The automation is now encapsulated in the triagebot-action, a standalone GitHub Action that allows other projects to adopt and adapt the system for their own needs, encouraging the creation of sustainable feedback loops that help maintainers focus on core development rather than backlog management.

Cloudflare has introduced a new Billable Usage API to enhance transparency in monitoring cloud expenses, catering to the needs of automation and programmatic usage tracking. This API provides a detailed breakdown of account usage and cost by product and service period for every usage-based Cloudflare service, including Workers, R2, D1, and more, allowing users to integrate this data with FinOps tools seamlessly. It aligns with the FinOps Open Cost and Usage Specification (FOCUS) to ensure familiarity for teams already using these standards, even though full conformance is still in progress. Cloudflare has partnered with Vantage, an infrastructure cost management platform, to integrate Cloudflare's cost data into a single view alongside other cloud providers, facilitating cross-provider allocation, anomaly detection, and cost management via FinOps agents. The API aims to provide real-time cost visibility, enabling finance teams and developers to efficiently track, forecast, and manage cloud expenses programmatically, with future improvements planned for finer-grained time windows and enterprise coverage.

Cloudflare introduces a preview of the @cloudflare/computer package, a novel agent runtime designed to optimize and scale the execution of code across various environments such as isolates, container sandboxes, and web browsers. This package aims to address the industry's increasing demand for efficient compute solutions by providing a shared and durable filesystem that agents can use to run tasks, select execution environments, and manage files. The @cloudflare/computer package leverages Cloudflare's long-standing innovations like Durable Objects and isolates, which allow for scalable, efficient, and cost-effective code execution. By offering a virtual filesystem with execution backends, it enables agents to perform tasks with minimal reliance on containers, thus balancing speed, capability, and cost. With the introduction of this open-source library, Cloudflare seeks to collaborate with developers to push the boundaries of agent scalability and efficiency, ultimately aiming to reduce container dependency for most coding and computational tasks.

Cloudflare has expanded its Workers RPC system, initially built on Cap'n Proto RPC, to enable seamless cross-language communication between JavaScript and Python Workers without requiring developers to define schemas or add dependencies. This system, which began as JavaScript-native RPC, allows for cross-language method calls, object sharing, and function returns, bridging the distinct type systems of JavaScript and Python through Pyodide's Foreign Function Interface (FFI). The integration ensures that developers can work as if they are using a single-language application, with Pyodide handling type translation and the workers-runtime-sdk ensuring native object handling across both environments. This advancement allows for the integration of Python packages in JavaScript applications, exemplified by using the Python package Pygments for syntax highlighting in a JavaScript Worker, showcasing the potential for building complex multi-language systems with minimal performance overhead.

Cloudflare is expanding its Workers platform to better support real-time voice AI applications and gRPC-based client-server communication by enabling inbound TCP connections and enhancing capabilities for gRPC applications. This development introduces a new handler, connect(socket), which allows Workers to accept inbound TCP sockets via Spectrum, Cloudflare's ingress proxy for non-HTTP traffic. This facilitates full-duplex, bidirectional gRPC communication from Cloudflare Containers, enabling low-latency interactions crucial for voice AI and colocated inference applications. Additionally, Cloudflare is offering gRPC-web support, allowing Workers to translate incoming and outgoing gRPC requests, thus enabling seamless integration with existing gRPC servers without requiring code changes from clients. This initiative is currently in private beta, reflecting Cloudflare's commitment to enhancing the versatility of its platform while maintaining performance and security features.

Workers AI leverages Cloudflare data centers to run inference for sophisticated open models like Moonshot's Kimi K-series and Z.ai's GLM, facing challenges due to their memory demands. To manage this, techniques such as quantizing the KV cache and compressing model weights are employed, which allow more efficient memory usage without compromising model accuracy. By storing the attention keys and values in lower precision and reducing model weights from 8-bit to 4-bit, significant reductions in memory footprint are achieved, enabling support for more concurrent requests and lowering operational costs. These optimizations are validated using SGLang, an open-source inference framework, ensuring the improvements are shared with the community. Additionally, an integrity checking system is implemented to safeguard shared memory caches, maintaining reliability with minimal performance impact. These strategies aim to enhance service capacity and cost efficiency while maintaining the quality of the models, reflecting an ongoing effort to optimize the deployment of high-demand AI models.

Agents Week explores the concept of an "Agent Cloud," reimagining cloud infrastructure to support the needs of autonomous agents rather than humans. Current cloud systems are optimized for human interaction, but agents require different considerations like speed, structure, and access. The initiative emphasizes developing a foundation built from the ground up for agents while also bridging the gap between existing human-centric systems and future agent-centric environments. Throughout the week, discussions will focus on understanding the necessary primitives and execution layers, updating the agentic software development lifecycle, and ensuring secure interactions between agents and organizations. Participants are encouraged to engage with their own agents to explore what they need from an Agent Cloud, fostering a collaborative approach to innovation in the field.