Cloudflare status: hosting issues and outage reports

Problems in the last 24 hours

The graph below depicts the number of Cloudflare reports received over the last 24 hours by time of day. When the number of reports exceeds the baseline, represented by the red line, an outage is determined.

Most Reported Problems

The following are the most recent problems reported by Cloudflare users through our website.

• Domains (33%)
• Cloud Services (29%)
• Web Tools (17%)
• Hosting (13%)
• E-mail (8%)

Live Outage Map

The most recent Cloudflare outage reports came from the following cities:

City	Problem Type	Report Time
Paris	Cloud Services	10 days ago
New York City	Hosting	13 days ago
Manchester	Domains	1 month ago
Angers	Cloud Services	1 month ago
London	Domains	2 months ago
Noida	Hosting	2 months ago

Full Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

Cloudflare Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Joshua Lipson (@Galkon) reported 3 minutes ago

  @dok2001 @Cloudflare @BraydenWilmoth still an issue btw. feels like CF can handle deleting a worker w/ over 100 deployments. instead it offloads the responsibility to me to use the API and paginate deletions in batches for ages to not hit rate limits

• 
  Quentin (@QuentLaBrute) reported 6 minutes ago

  @Sousinr Hope they never read this but cloudflare feels like a blessing, way too cheap for everything you get to do with it! they’re a critical part of my stack and have never paid them anything

• 
  Akos (@akoskm) reported 6 minutes ago

  @martindonadieu is it through the remote session? like cursor running on your mac and usng remote contorl on your phone? I'm not really into the cloud version, maybe I'm not using it properly, but it can't: - read GitHub issues from my private repos, even tho the Cursor app is installed and has permission to read them - i keep my cloudflare etc secrets locally in an .env file, cloud agent doesn't have access to that - i develop ios apps and the e2e ui tests need an actual macos running what am i doing wrong? appreciate the answer!

• 
  Toolport (@toolportapp) reported 7 minutes ago

  Cloudflare's CTO, Dane Knecht, emailed their entire customer list about running every AI agent through MCP portals, with scoped tool access and a Code Mode they say cuts token costs ~93%. We build Toolport. This is our exact category. A few notes from the small end of the pond. 1. The announcement is two products wearing one trenchcoat. AI Gateway is a model-layer proxy. It routes your OpenAI/Anthropic calls and meters model spend, same lane as OpenRouter or LiteLLM. MCP portals are the tool layer, deciding which servers and tools each agent can touch. Different problems, different failure modes. If you're evaluating either one, don't let the bundle blur that line. 2. Their ~93% Code Mode number lands in the same range we've been measuring with lazy discovery for the past three months. Loading dozens of tool schemas into context on every request was always the wrong default. It's nice that it's no longer just indie gateway builders saying so. 3. The structural difference. Their portals require being a Cloudflare customer and routing agent traffic through their edge. Toolport runs on your own machine and works with Claude Desktop, Cursor, and 25 other clients, with no vendor in the traffic path. When the traffic is your prompts, your credentials, and your tool outputs, where it flows is not a detail. 4. Scoped access is becoming table stakes. The harder question is whether the tool you approved yesterday is the same tool running today. Toolport pins tool integrity, quarantines drift, and scans for injection. A DLP scan on the wire doesn't catch a server that changed underneath you. Watching a giant walk into your category is a strange feeling. It's also the strongest signal yet that this layer needs to exist. We just think it should live on your machine, not in someone else's cloud.

• 
  Darren ♿️🎮📸🎧📀 Richardson (@S0UNDK1LLAH66) reported 8 minutes ago

  @MrBonMot Thing is, Jim. Is that as much as we become reliant on the web? If Cloudflare goes down or a major outage - we are stuck in transactional hell. Cash is king..always.

• 
  Freddyツ (@InvestorFreddy1) reported 9 minutes ago

  In my days of bull-posting $NBIS, I haven't yet made a post about the other great companies that $NBIS is a part of. This is not a holding company. It is a portfolio of bets on infrastructure that the next decade of AI physically cannot run without - and most people covering it have not done the work to understand what is actually inside. ClickHouse is the one that keeps me up at night in the best way. Real-time analytics at a scale that makes most databases look like Excel. Cloudflare runs on it. Spotify runs on it. The entire AI observability space - every company trying to monitor model behavior in production - is converging on columnar databases exactly like this. ClickHouse is not a niche tool anymore. It is quietly becoming the plumbing. Then there is Avride. Autonomous delivery and robotaxi software, spun out of Yandex. They have active pilots with Grubhub and are deploying sidewalk robots in real cities today - not a demo, not a proof of concept, actual commercial operations. The market keeps pricing autonomous as a 2030 story. Avride is operating in 2026. CoreWeave exposure through the stack ties the whole thesis together. When GPU demand spikes, the companies that manage, route, and analyze that compute workload scale with it. NBIS is not one bet on AI. It is six or seven bets on the infrastructure layer sitting just below the headline names everyone already owns. Today's drop is a gift or a warning. I think it is a gift because none of the underlying businesses changed today - only the price did. What I am watching is the ClickHouse commercialization timeline. If they move toward a direct enterprise licensing push or any IPO signal before end of year, the market has not remotely priced that in yet. That re-rating would be violent. I bought at $50 and I have not sold a single share. That should tell you where I stand. What is the one NBIS holding you think is most underappreciated right now? #NBIS #AI

• 
  Dr. Xi Zeng (@xiz25) reported 9 minutes ago

  @dhh @Cloudflare Putting speedtest and DNS choice in the same panel is the detail: diagnosis and action live together. Most network UIs show status, then make users hunt elsewhere for control. Which metric actually changed your behavior after using it?

• 
  DFIR Radar (@DFIR_Radar) reported 13 minutes ago

  Chaos ransomware's msaRAT routes C2 entirely through Chrome or Edge via Chrome DevTools Protocol, making attacker traffic invisible on the wire. The RAT process never touches the network directly. - msaRAT is a Rust-based RAT attributed to the Chaos ransomware group. Initial access delivers update_ms.msi via curl over HTTP on port 443, bypassing port-based firewall rules. The MSI impersonates a Windows update and loads the RAT DLL directly into memory, leaving no obvious on-disk payload to catch. - C2 is split across two legitimate services: Cloudflare Workers at is-01-ast.ols-img-12.workers[.]dev handles SDP signaling to establish a WebRTC channel, then drops out entirely. All subsequent commands flow over a WebRTC DataChannel relayed through Twilio, so the attacker's real server IP never appears in traffic. - The RAT binds only to 127.0.0[.]1. All external traffic originates from the browser process itself, launched headless with remote debugging enabled. CSP is bypassed via a CDP command and JavaScript injected into the browser handles every network operation. What a defender sees is a browser making HTTPS calls to Cloudflare and WebRTC to Twilio, both normal. - Talos found msaRAT deployed between initial access and the Chaos encryptor, confirming it serves pre-encryption reconnaissance. Double encryption: DTLS from WebRTC plus ChaCha-Poly1305 over an ECDH-derived key means stripping DTLS still yields ciphertext. #DFIR_Radar

• 
  Jerry Combs (@PointBlueTech) reported 16 minutes ago

  @PH0ENIXSMITH @thdxr Exactly. I never have to do anything in CloudFlare myself.

• 
  Michael Timbs (@michael_timbs) reported 19 minutes ago

  @thdxr Hahahahha. Improbable. One of the many reasons there’s very serious applications deployed on Cloudflare. Primitives are just wrong with terrible APIs the entire way down

• 
  Rishi Raj Jain (@rishi_raj_jain_) reported 21 minutes ago

  @SantoshYadavDev @astrodotbuild @Cloudflare CF Pages are basically replaced with Worker. Lmk if you need any help!

• 
  Muhammad Tahir (@Ta56980T) reported 22 minutes ago

  @Cloudflare we have to stop solving problems for sake of problems, but focus on how to Never let those problems even happen (find out of box solution that why it even happens? how to make such technologies which are resilient to such attacks by their architecture.

• 
  Voxcadax (@CovaDax) reported 27 minutes ago

  @bananajoexbt @ivvanex @PlayStation Even if it's not our fault. Oops cloudflare broke, sorry, talk to them, no, cloudflare problem is our problem now.

• 
  Harman (@itsharmanjot) reported 28 minutes ago

  Most developers scan file extensions and call it security. Attackers don’t care what your file extension check says. There’s an open-source tool that actually scans the bytes before a file ever touches your disk. It’s called pompelmi. A TypeScript toolkit for Node.js that scans uploads in-process, MIME spoofing, ZIP bombs, nested archives, polyglot files, before storage, before your CDN spreads it, before it becomes someone else’s problem. → Scans in-process with zero outbound network calls, bytes never leave your server, no cloud API, no daemon required → Deep ZIP inspection catches nesting depth, file count, and archive traversal tricks that simple extension checks miss entirely → Optional ClamAV antivirus integration plus optional YARA rule support, for teams that want signature-based detection layered on top → Drop-in adapters for Express, Fastify, Koa, NestJS, and Next.js, plus support for Bun, Deno, and Cloudflare Workers → Built-in security scorecard grades your actual upload configuration A through F, so you know your real exposure, not just whether the tool is installed → Quarantine mode auto-moves flagged files with a sidecar JSON report, so a suspicious upload gets isolated automatically instead of silently landing in your storage bucket Any app that accepts user uploads is dealing with a real attack surface most teams treat as an afterthought. pompelmi’s whole design is catching the risk at the entry point, not after the file’s already spread across your systems. 337 GitHub stars.

• 
  Lord Bean (@lifeisameeme) reported 32 minutes ago

  Deployed a static site on Cloudflare Pages last week. Free tier, 𝘂𝗻𝗹𝗶𝗺𝗶𝘁𝗲𝗱 𝗯𝗮𝗻𝗱𝘄𝗶𝗱𝘁𝗵, global CDN. Then spent 3 hours on gotchas nobody writes down. → A project made via `wrangler pages deploy` is 𝗗𝗶𝗿𝗲𝗰𝘁 𝗨𝗽𝗹𝗼𝗮𝗱 𝗳𝗼𝗿𝗲𝘃𝗲𝗿: no "connect to ***" option ever shows up for it → The official GitHub Action pins an old wrangler that breaks with Pages-scoped tokens. Fix: force `wranglerVersion: "4.98.0"` → Cache headers are immutable for a year. Swap icon.png for new bytes under the same name, and the CDN just keeps serving the old file Not gonna lie, that last one cost me the most time. Version your filenames. 🔗 Docs in the reply 👇

• 
  Synapse Brief (@Synapse_Brief) reported 32 minutes ago

  Containers were never built for this. Cloudflare just proved it. On July 14, Cloudflare shipped Dynamic Worker Loader, an open beta feature that lets a Worker spin up a new Worker at runtime running AI-generated code. No container spin-up. No hundreds of milliseconds of cold start. Just an isolate booting in a few milliseconds using a few megabytes of memory. Cloudflare's own numbers: isolates start in single-digit milliseconds and sip a few MB of RAM. A typical container takes hundreds of milliseconds and eats hundreds of MB. Cloudflare is calling this roughly 100x faster and up to 100x more memory efficient. That's their claim, not an independent benchmark — nobody's third-party tested it yet. But even half true, it changes the math on running untrusted, AI-written code at scale. Here's the part that actually matters more than the speed number. This isn't just a faster sandbox. It's Cloudflare betting that agents are done being chatbots that make sequential tool calls, and are becoming code runners instead. Their framing: the LLM writes TypeScript, that code executes inside a Dynamic Worker, and only the result comes back into context. No fifteen round-trip tool calls burning tokens. One execution, one output. They're also pushing TypeScript RPC as the interface layer over raw HTTP for agent tooling — shorter, more precise, easier to constrain than a pile of REST endpoints. If that thesis holds, TypeScript quietly becomes the default control plane for how agents actually do things, not just talk about doing things. Security-wise, it's locked down by default: no outbound network access unless you explicitly wire it up through globalOutbound, with interception and credential injection available if you need it. Runs across every one of Cloudflare's hundreds of edge locations, usually on the same machine and thread as the parent Worker. And the pricing is almost a joke next to what inference costs: $0.002 per unique Worker loaded per day, waived entirely during beta. What's missing: real customer traction numbers for this specific feature, and no outside party has stress-tested the 100x claim across varied workloads. Take the performance number as Cloudflare's marketing until someone independently reproduces it. Still — pair this with Cloudflare shipping temporary accounts for agents in the same window, and the pattern is obvious. They're not building a feature. They're building the execution layer underneath every AI agent that needs to actually do things instead of just chat about them. Curious if anyone's already load-tested this in production, or if we're all still just reading the launch post.

• 
  Ara T. Howard (@drawohara) reported 34 minutes ago

  today, with the help of 3 agents, I managed to create a cloudflare API token JFC

• 
  Publisher in a Box (@publisherinabox) reported 34 minutes ago

  Click-through on the number one organic result drops about 58% where an AI Overview appears. @ahrefs measured it across 300,000 keywords. That number sits inside a data-heavy breakdown of where AI search stands right now. Five points matter most for operators: 1. Citations alone are not the whole picture. Wil Reynolds tracked a 1,900% month-over-month jump in ChatGPT citations to a single page on his site. Business impact: close to zero. Counting how many times an AI engine mentions you is half a metric. It does not tell you whether you were recommended or just named, whether the description was accurate, or whether it happened on the engine your buyers actually use. 2. Google's own reporting confirms the gap. The new AI performance report in Search Console shows impressions only. No clicks, no queries, no CTR. Marketers asked for attribution data and got a vanity counter. You cannot make a business decision on impressions alone. 3. The click loss is quantified, not anecdotal. The 58% is measured click-through on position one across 300,000 keywords. Publishers who depend entirely on organic search clicks are watching a measurable decline, not an imagined one. 4. Most agentic infrastructure is not being read. Ahrefs studied 137,000 sites and found 97% of llms.txt files never get read by a crawler. Meanwhile Cloudflare reported in June 2026 that bot and agent traffic crossed 50% of all internet traffic for the first time. Technical retrievability is the whole game here. Schema, llms.txt, crawler access. It is the gap that closes fastest and the one almost nobody has verified actually works. 5. GEO demand is climbing from a standing start. "Generative engine optimization" now pulls 7,900 US searches a month, up 997%. The entire AI attribution keyword cluster is up 430% year over year. Operators are searching for this because they feel the traffic shift and want a playbook. Here is the part most people skip. GEO gets you cited inside the AI answer, SEO gets you ranked in a list of links. Four engines answer differently, ChatGPT, Claude, Perplexity and Gemini, and most publishers appear on one and are invisible on the other three. Sentiment counts as much as volume, because being named is not the same as being recommended. We built our GEO scoring before most teams knew the term, and we ran it on our own properties before we ran it for anyone else. The AI referral traffic followed. That is the only reason we are confident enough to publish the methodology. For publishers building a real business, GEO is another revenue line sitting alongside Facebook, the website with Google Discover, and the newsletter. Facebook is the base the rest of it is built on. It pays publishers directly and sends readers to the site, and that is what funds every other channel. A diversified publishing business treats each discovery surface as a channel worth measuring on its own. The operators who figure out AI referral traffic now are claiming ground the rest of the field will have to buy back later. The ground in your niche is still open. How we measure it and how you claim it, in the first reply 👇

• 
  chrißy (@chribdotnet) reported 35 minutes ago

  i need to set up cloudflare today and honestly cloudflare scares ******** out of me if i could make my own cloudflare i surely would but i just learned about /POST last night all this **** is easier than i expected ngl but still hard

• 
  Sumanth (@Sumanth_077) reported 36 minutes ago

  Turn any website into agent-ready data! Loop engineering is about designing systems that run agents autonomously. Instead of prompting your agent manually each turn, you write a loop that finds the work, hands it to the agent, checks what came back, and decides what happens next. Your job is to design the loop once and walk away. But loops that need live information from the web hit a real constraint. JS-heavy pages return empty content. Anti-bot systems return challenge pages. Login walls block access entirely. When the model gets weak context back, it still responds - just less accurately. Anakin is building the source-access layer underneath agents. URL Scraper turns any URL into clean Markdown, HTML, or structured content immediately usable by an LLM. Built for scale across 200M+ active websites globally, including a large chunk of Cloudflare and Akamai-protected pages. Authenticated sessions handle content behind login walls. Wire handles workflow-heavy sites. Login flows, navigation, form submission, report exports - all accessible through a stable API. Define the workflow once and Wire keeps it working as websites change. Key capabilities: • Clean Markdown, HTML, or structured output from 200M+ active websites • Built for difficult pages including Cloudflare and Akamai-protected sources • Authenticated sessions for content behind login walls • Wire for login flows, navigation, form submission, and export-based access • Useful for AI agents, RAG systems, finance intelligence, and vertical AI workflows I've shared the link in the replies!

• 
  Milk Road AI (@MilkRoadAI) reported 39 minutes ago

  Open-source is dying and the companies that survive it are about to get very rich (Save this). That's the uncomfortable truth in Dylan Patel's take, American open source AI is basically dead. Meta has gone quiet on Llama, Mistral, once the loudest open source advocate in the West, shifted its flagship models to proprietary licensing while charging five to ten times more than comparable Chinese models for similar performance. So the only frontier level open models left are Chinese, Qwen, DeepSeek, Kimi, GLM and the labs building them barely profit from giving them away. The money instead flows downstream, to whoever hosts, serves and charges for access to those free weights. Qwen overtook Meta's Llama as the most downloaded model family on Hugging Face in 2026, and Chinese open models now out download American ones globally by a wide margin. Kimi K3 was ranked the top open source model in the world by LMArena. None of that revenue lands with the Chinese labs themselves but rather lands with the inference layer running on top. Inference is already the biggest chunk of the AI compute market, with cloud inference alone estimated near $50 billion in 2026 and growing around 60% a year, dwarfing training infrastructure spend. Token pricing has collapsed roughly a thousandfold over three years which sounds bad for margins until volume growth outpaces the price decline, keeping total inference revenue climbing. Mistral's own pivot away from open weights shows what happens when a lab tries to charge premium prices in a market Chinese competitors are commoditizing its newest model is losing on both cost and quality to rivals a fraction of the price. If Chinese labs eventually decide there's no financial reason to keep releasing frontier models for free, the open-source pipeline could dry up overnight. However, here are the publicly traded infrastructure plays positioned to benefit if open source continues to dominate. Nebius provides the same raw compute layer underneath inference demand, without needing to bet on any single model's survival. AMD is chasing that same inference chip opportunity with its MI series accelerators, positioning itself as the main alternative supplier once inference volume keeps compounding. Cloudflare (NET) benefits through its Workers AI platform, which increasingly serves as the delivery layer pushing open weight models out to edge devices and apps cheaply. Microsoft (MSFT), Amazon (AMZN), and Alphabet (GOOGL) all benefit as the hyperscalers whose cloud platforms host the bulk of enterprise inference workloads, collecting compute revenue no matter which model an enterprise ultimately runs. Milk Road Pro is tracking all the biggest beneficiaries of open source AI, if you want access to all our AI trades around this trend, you can come join us for just $1 using the link below!

• 
  Colin Son (@txmedai) reported 40 minutes ago

  @ptremblay Yes Cloudflare is very good. And free tier is amazingly generous. Isnt quiet idiot proof (why do I have independent turnon AI gateway or R2 or all these individual features) but at least there’s not a million layers of IAM. GCP and AWS are fine

• 
  Matt Schober (@migratewithmatt) reported 43 minutes ago

  Stopping the bad guys with Cloudflare: 2,176 malicious requests blocked or challenged in the last month #cloudflare

• 
  Pavel Shu (@shu_pavel) reported 45 minutes ago

  It’s surprising how many people in our industry aren’t even trying. To them, AI is just a threat. Maybe they’re right. If not today, then in a year any mechanic will be able to build whatever they need. No APIs. No Cloudflare. No databases. But damn, it’s exciting to finally stop pushing pixels and wrestling with state sync in yet another React “prototype” just because you don’t know backend. It feels like pure ******* creativity.

• 
  Jonas Templestein (@jonas) reported 46 minutes ago

  @thdxr (Not a serious application) Alchemy is v good but unfortunately cloudflare is investing heavily in wrangler for local dev and it sucks not being able to run apps locally So we made a little script that takes env config and secrets in and produces wrangler files and wrangler commands It works quite well for us now - but it does feel like we are building half of IaC because we need to sequence resource creation and tear down and deal with flaky control plane APIs etc

• 
  Trishool | SN23 (@trishoolai) reported 49 minutes ago

  AI is at the same inflection point web security hit twenty years ago. The attacks are real, the damage is mounting, and the industry is starting to realise that shipping without a safety layer isn't a risk worth taking. Cloudflare didn't create that shift in web security. It became one of the companies that defined it. We believe the same thing is happening in AI right now. AI agents have deleted production databases, exposed sensitive customer data, and taken actions their creators never intended. As AI systems do more and touch more, the cost of getting safety wrong keeps growing. The internet eventually stopped expecting every company to build its own security infrastructure because the problem outgrew what individual teams could manage. Shared infrastructure became the better answer. AI safety is heading in the same direction. That's the future we're building toward with HaloGuard on Bittensor. Production-ready, peer-reviewed, open weights, and built to protect AI systems across 46 languages, backed by a decentralised network that gets stronger every week. AI safety isn't a problem that gets solved once. It's infrastructure that has to keep learning as the threat landscape evolves. That is what we are building on Bittensor.

• 
  nik skld (@nikskld) reported 50 minutes ago

  ANTHROPIC SHIPPED A CRON SCHEDULER FOR AI AGENTS AND ALMOST NOBODY COVERED IT. Every thread you’ve seen about Claude Managed Agents is still recycling the April 8 launch post. Three months later the product barely resembles that announcement. What actually shipped since: •Scheduled deployments (June 9). Give an agent a cron schedule. Every time it fires, the agent opens a fresh session and completes the task. No scheduler to build, no scheduler to host. •Environment variables in vaults. Register an API key with the domains it’s allowed to reach. The sandbox only holds a placeholder. The real key gets attached at the network boundary, on approved domains only. The model never sees it. •Self-hosted sandboxes (May 19). The agent loop stays on Anthropic’s infrastructure. Tool execution moves inside your perimeter. Cloudflare, Daytona, Modal and Vercel are supported out of the box. •Memory graduated from research preview to public beta. Only MCP tunnels and dreaming are still gated behind an access request. Pricing almost nobody quotes correctly: standard token rates plus $0.08 per session-hour. Runtime is metered to the millisecond and only accrues while the session status is “running.” Waiting on your reply or a tool confirmation costs nothing. Now the part the hype threads skip. Managed Agents is stateful by design. Conversation history, sandbox state and outputs are stored server-side. Anthropic’s own docs state that this makes it ineligible for Zero Data Retention and for HIPAA BAA coverage. So every “finally, a secure alternative to hosting agents yourself” take has it exactly backwards. You get sandboxing, scoped permissions and execution tracing. You give up data retention guarantees. For legal, health and financial workloads that trade is the whole decision. Old way: build the loop, the sandbox, the scheduler and the secret vault yourself, then ship in months. New way: define the agent, point it at a cron, ship in days, and read the compliance page before you move anything sensitive through it.

• 
  Gavin Keulks (@GavinKeulks) reported 51 minutes ago

  @DrDavidVernon @britishlibrary Thank you for those kind words. The site was severely damaged by a server upgrade at my university, and for years I thought it was gone forever. Thankfully, repair (albeit tedious) has been possible. Now I've taken full ownership of all aspects of the site, so if it's damaged again, it will be because Cloudflare accidentally shut down the internet again (as it did a few years ago).

• 
  Chris Board (@chrisboard_) reported 56 minutes ago

  From a network trace, it looks like a TCP handshake completes, but then a ClientHello is sent, but no response is received so keeps getting retransmitted until the timeout happens. I know there's nothing being shown on cloudflare status but this does seem like it might be related to cloudflare.

• 
  Matt (@meszmatew) reported 58 minutes ago

  Is anyone else having issues with cloudflare billing?