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.

• Cloud Services (43%)
• Domains (21%)
• Web Tools (14%)
• Hosting (14%)
• E-mail (7%)

Live Outage Map

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

City	Problem Type	Report Time
New York City	Cloud Services	17 days ago
Los Angeles	Cloud Services	18 days ago
Paris	Cloud Services	1 month ago
New York City	Hosting	1 month ago
Manchester	Domains	2 months ago
Angers	Cloud Services	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:

• 
  🐗 Giacomo Balli (@BigBalli) reported 5 minutes ago

  Woke up to a saturated PHP-FPM pool. Overnight someone pulled 16,420 requests off one of my sites in seven hours, from 9,509 different IPs. All of it China Mobile CMNET. 36.128.0.0/10 and 36.212.0.0/14, spread across 18 /16 blocks. The detail that took me longest to see: 15,628 distinct URLs. That's 1.05 requests per URL. It walked the entire URL space, one fetch each. Every request was a cold miss at the edge by construction. Cloudflare cached all of it correctly and absorbed none of it. Seven rotating Chrome user agents, every one a real browser string, so all my UA-based rules sailed right past it. No reverse DNS. No self-identification. It never asked for robots.txt. It averaged 1.7 requests per IP, which is below every per-IP threshold I had. Zero requests hit the rate limiter. Blocked it by netblock in the end. Fifteen days of logs show 43 requests total from those ranges to my 40-odd other sites, so the collateral is nothing. Anyone else getting hammered out of CMNET lately? Trying to work out whether this is pointed at me or spraying everyone.

• 
  Rick A.F. (@FoxRick01) reported 5 minutes ago

  @andreigec My agents connected to Cloudflare handle all this admin, it’s not an issue at all tbh

• 
  Urjit (@urjit_) reported 5 minutes ago

  @tannerlinsley @tan_stack next on the todo for today after migrating from cloudflare project think to my own harness, and it didnt even support tanstack ai anyways 😢

• 
  assertchris (@assertchris) reported 6 minutes ago

  I thought the 10 custom domain limit (on the Laravel Cloud plan I'm on) would be an issue. It just pushed me to handle more things directly in CloudFlare. A much better solution of those things that need mail DNS and a simple redirect.

• 
  Choblin (@choblin29) reported 7 minutes ago

  Greg Brockman says GPT-5.6 Sol found 13 security issues on his personal website in about 15 minutes then ChatGPT Work spent roughly an hour fixing them -- changing DNS, TLS, removing jQuery and moving the site to Cloudflare Pages.

• 
  Tobias_Petry.sql (@tobias_petry) reported 9 minutes ago

  @scheemunai @shanejones The number of people disabling scripts is neglible, mostly non-existent. The bigger problem will be adblockers blocking the Google Analytics code. So your real number I guess will be say like ~20% higher than GA reports. The Cloudflare one is highly inflated all the time.

• 
  Victor Garcia (@vmg__0) reported 11 minutes ago

  @kellabyte Cloudflare doesn't give the details on how they do durability. For CellD I think you can have different durability levels, like making every write be RPO 0 but I think it's configurable to have better write performance at the expense of there being potential data loss if the node goes kapoot. I think I see what you meant about the issue with reads after writes because if the local sqlite db is lost and there was a read after a write and the write isn't durable the client could see some inconsistent state.

• 
  Tony Simons (@tonysimons_) reported 20 minutes ago

  @PhantomByteAI My first thought was Cloudflare issue, but since I can’t connect via Telegram even, I’m thinking this one is on my end. I did have it doing some video work, so I wonder if it tanked resources and just froze up.

• 
  Corey J. Gallon (@CoreyGallon) reported 22 minutes ago

  Personal apps break the cloud architecture we've spent 25 years building. That's the single point @KentonVarda makes in "Gadgets: Personal app vibe coding that is actually safe," on @aiDotEngineer's YouTube. Kenton is a Principal Engineer at Cloudflare and started the Workers project in 2017. The talk walks through a working platform he built to test the idea, and it's specific about the sandboxing that makes user-modified code safe to run. - The plugin-system death spiral. A developer drowning in one-off feature requests decides to rewrite around plugins, the rewrite never ships, and neither do the features. - The alternative is users editing their own copy. The developer ships a clean core app, and anyone who needs a feature asks an agent to add it, just for them. - Server-per-user is the blocker. One blessed version of an app running on your server is convenient for developers and makes customization impossible, which is exactly what today's vibe coding platforms are built on top of. - Gadgets work like documents, not deployments. Think Google Docs: hundreds of gadgets, each one an app with its own code, each one shareable. - Sharing lives in the platform, not the app. Because a gadget is a single shareable thing, access control is implemented underneath it, so the app can't get it wrong. - Blueprints are code without data. Export a gadget you like as a blueprint, and other people instantiate their own gadget from it. - The agent modifies the app, not just the content. Asked to build a slide deck, Claude added strikethrough, text centering, and an SVG paste box to the Slides app itself when the features it needed weren't there. - Security by containment, not by correct code. The client runs in a null-origin iframe sandbox under CSP that can only postMessage to the parent; the server runs in a dynamic worker sandbox. Neither can reach the outside world, so an XSS bug leaks nothing. - Cap'n Web RPC connects the two halves. The postMessage channel carries an RPC session through to the gadget's server code, written as a durable object. - No containers, no database. The whole thing runs on dynamic workers and durable objects, and the entire demo ran locally on his laptop on workerd, the open source Workers runtime. He also explains why the code isn't on GitHub yet, which he'd promised in the abstract. I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!

• 
  Erik Langille (@eriklangille) reported 22 minutes ago

  @BraydenWilmoth can you fix the two terrible iOS apps cloudflare already has? (1.1.1.1 and Cloudflare One) Tried using Warp for CF Zero Trust and it kept erroring. Turns out I needed this other CF One app. Took like 2 hours of debugging Tailscale is way smoother

• 
  Antonio Gomes (@onchainantonio) reported 22 minutes ago

  @Googlecloud's recent post-quantum roadmap highlights just how complex this migration is. The most honest line in it is about hardware. Google says some physical replacement cycles will run past 2029. This is a company that owns its data centres, its silicon, its browser and its certificate authority. And it still cannot promise the hardware layer is done by the end of the decade. That is the whole migration in one sentence. Post-quantum is a systems programme, not an algorithm swap. Three things in the roadmap that make the point: → Key exchange was the easy part. Hybrid ML-KEM is already live on Google's endpoints. Encryption in transit moves fast because it only needs two parties to agree. → Signatures are the hard part. ML-DSA and SLH-DSA are too large for the existing WebPKI to carry, so Google is testing Merkle Tree Certificates with Chrome and Cloudflare. They did not swap the signature. They redesigned how trust is distributed. → Roots of trust are the slow part. Caliptra, TPM 2.0, OpenTitan for quantum-secure boot. Silicon has a shipping cycle. No software update shortens it. Now remove Google's advantages. No single owner of the stack. No control over the clients. Keys and signatures published permanently the day they were made. That is what a decentralised network is migrating from. It is why "we'll upgrade later" is a much bigger claim there than it sounds. The organisations with the most control are being the most careful about the timeline. Worth sitting with.

• 
  SK64 (@stonkeykong64) reported 23 minutes ago

  Is Cloudflare main site down or what?

• 
  Eric Sauvageau (@RMerlinDev) reported 25 minutes ago

  For the curious, I already have a plan B if things ever went south with @Github . For years I've had a gitweb mirror running on a VPS, secured behind a Cloudflare App login. Since I don't care about any of the CI/agent stuff, I would only miss PR and Wiki capabilities.

• 
  Kelano (@kelanoo) reported 27 minutes ago

  5 free repos just blew up on GitHub this week, and they all upgrade your AI coding setup. 1. Agent Skills, 87,000 stars. It's eight commands that cover your whole dev cycle, spec, plan, build, test, review, ship, so your agent stops skipping steps. 2. book-to-skill. Point it at any book or PDF and it turns the whole thing into a skill, so your AI stops guessing at best practices it never actually read. 3. Cloudflare Computer. It gives your agent an actual Linux computer, so it never loses your work the second a session ends. 4. TencentDB Agent Memory. one shared memory for your whole team of agents, so you never explain the same thing five times. 5. authentik. a free open source auth server with 24,000 stars, so you can stop paying Auth0 and Clerk just to put a login on your app.

• 
  DNLaunch (@dnlaunch) reported 28 minutes ago

  @TruDomainz Google Search Console and CloudFlare can help

• 
  Ido Bronstein (@IdoBronstein) reported 30 minutes ago

  On August 6th, @Cloudflare launched a web browser no human will ever use. It's called Kitesurf, and it's built only for AI agents. Instead of forking Chromium — the engine behind almost every browser automation tool of the past decade — Cloudflare put together a new lightweight engine in Rust. In their own tests it uses up to 3.8x less CPU and 7x less memory than Chromium, at the cost of running a bit slower per task (@TechCrunch). Read past the specs. The cost of letting an agent act on the web just dropped by a large multiple. When something gets that much cheaper, you don't get a little more of it. You get a flood. That's where I'd slow down. Every company is about to run far more agents, doing far more things, for far less money. The bottleneck stops being "can we afford to run agents" and becomes "can we trust what they do at that volume." Here's the part that changes. A wrong answer used to be one bad report that one person caught. At agent speed, one wrong definition becomes a thousand wrong actions before anyone looks. The mistake isn't worse. There are just many more copies of it, made much faster. So before you scale the agents, fix what they read. Clear definitions. One source of truth. A named owner for each definition, so when something breaks you know who fixes it. Cheap agents multiply whatever you feed them. Make sure it's worth multiplying.

• 
  Milan Jovanović (@mjovanovictech) reported 33 minutes ago

  How do you handle errors coming from external APIs? Here's an example using the Keycloak REST API. If you're unfamiliar with it, Keycloak is an OAuth-compliant identity server. There's an API endpoint you can call to register a new user with Keycloak. If the provided email isn't unique, you'll get an error (409). The correct response code for this type of error is 409 Conflict. You are trying to create a "resource" that already exists. What would you do here? Since I know how to handle this exception, I would stop the execution here and return an error message to the caller. I approach errors and exceptions differently in my code. Another one I've had to deal with recently is hitting my usage limits with Cloudflare Email Sending (429). Nothing I could do about this, so I just decided to log the exception and fail fast.

• 
  ryqwzr (@ryqwzrbuilds) reported 34 minutes ago

  The payment layer is where this gets interesting. Cloudflare's pay-per-crawl beta uses HTTP 402, the old `Payment Required` status code, so publishers can allow, block, or charge crawlers. Cloudflare also says sites on its network already send over a billion 402 responses per day to bots and crawlers. The signal exists. The settlement layer is the missing piece.

• 
  Mark Statkus (@cybercussion) reported 34 minutes ago

  @silvestru @wholemars I'm similarly at $200/mo USD but its lasting 5 days the last month or so (least not til they stop the 50% credit deal they have). General tips - - Code Abstractions confuse agents and cost time and tokens. I'd ask your agents not just what they are doing, but why is this costing me 180K tokens. - Do you have something blowing out your context window (too many MCP, Skills, MD files etc) ? - Agents can also reason over your session logs, that can give you some idea where the churn is. Tool use, lots of grepping and searching a large project etc. - *** though fine for humans, is inefficient to agents. And humans are papering that over right now -- I ran from this in April. I'm not saying don't use ***, I'm just saying its not going to scale for where we are going. I generally call this a coordination tax or ceremony tax I no longer wish to pay. i.e. I don't have to keep my SDD separated or apart and I can in tandom work with Grok, Claude, Codex. They even find each other on the substate which is entertaining. They work together, take on parts of a plan and review each other's work. - IDE's also index code bases. You can play that forward by having a agent run ahead on project to lower the re-reasoning over the project as every cold start you'll start off doing it all over again. (this is a common complaint among coworkers working on old legacy projects that pre-date AI). And then the trick is getting agents to reach for it instead of ls, grep, cat etc... I have roughly done new VCS in Rust (context/intent/incremental memory), Session Reasoning, Project Indexing, Agent Chat across fleet which leads to Peer level work (new and interesting behaviors you'll never gain without it), Vendor based CLI Tool model agnostic used for implementation/review in any harness and then on the infra side I collapsed DevOps into runnable markdown with state. Agents can work over this more cleanly, find gaps, solutions, note them faster than me having a agent whip up Terraform or something else. And then you're essentially able to dogfood taking these to AWS, Cloudflare, GCP, Azure etc and even do cost analysis etc. Goal Speed boosts, lower token use.

• 
  Chong-U (@chongdashu) reported 38 minutes ago

  Hey @Cloudflare I’ve had an open ticket for domains purchased on Cloudflare that never went through ( i had to get them on an alternate provider) But it’s been a month with no reply and i got invoiced for the domains via Cloudflare . Some help please?

• 
  XXI (@XXIfomo) reported 41 minutes ago

  A 21-YEAR-OLD IN COPENHAGEN BUILT A HEX-GRID LIGHT PUZZLE IN ELEVEN DAYS AND HIT $4,200 A MONTH BY WEEK EIGHT. HE SPENT $205 TO SHIP IT. MOST PEOPLE STILL HAVEN'T SHIPPED ANYTHING. He is 21. Solo. No studio. He works from a shared desk in Nørrebro. Claude Code writes the puzzle generator, the reflection math, and the leaderboard. He hand-tunes the visual polish and refuses to let the model touch the lighting effects. The whole game runs on Cloudflare Workers for five dollars a month. The player never asks who built it. She screenshots her solution and asks the internet how she pulled it off in seven moves. Everyone else spends three months vibe-coding an original masterpiece that nobody plays. He picked a shape people already knew how to solve, added a limited-moves counter, and dropped the link in three subreddits on a Tuesday. By Friday it had 340 daily players. By week seven it had 47,000. Build time = 11 days Total cost to ship = $205 Claude Code Max subscription = $200 per month Hosting = $5 per month Daily active users at week eight = 47,000 Monthly recurring revenue = $4,200 Marketing spend = $0 For nine days nothing worked. The generator produced puzzles with no valid solutions. He almost killed the project on day ten. On day eleven he switched the generator to reverse-solve from a target state and every level shipped playable. The first hundred players arrived before breakfast. Monument Valley shipped in three years with a team of ten. The kid in Copenhagen shipped a shape people already knew how to solve in eleven days. bookmark this and read the article below

• 
  RED BARN NEWS 🇺🇸 (@JasperRoads) reported 46 minutes ago

  1/2 Quick tip that can save your parents/grandparents a ton of headaches: Grandma and Grandad keep getting hit with phishing texts and links. Every few months the phone gets completely wrecked and needs a factory reset. Changing the DNS to Cloudflare Family (1.1.1.3 + 1.0.0.3) blocks known phishing, malware, and scam sites at the DNS level. When they click a bad link, the page just doesn’t load instead of infecting the phone. It’s free, doesn’t slow anything down, and works quietly in the background. Not 100% perfect, but it stops a huge chunk of the common phishing junk older people get targeted with.

• 
  Aman (@devjedi404) reported 48 minutes ago

  Update on my Hot Wheels bot: It has officially gone from "random script I'm hoping works" to a proper monitoring pipeline 😭 Also caught a nasty false-positive bug where an ADD button from a recommended product could be mistaken for my wishlisted car. Persistent browser, saved location, concurrent wishlist checks, crash recovery and Cloudflare backoffs are in too. Now I just need Blinkit to actually restock the damn cars.

• 
  Vivek Maskara (@maskaravivek) reported 52 minutes ago

  1. Vercel → Cloudflare + AWS Amplify I had around 10 projects on Vercel Pro, andthe fixed Pro cost plus builds and runtime usage was costing me a lot. I moved the static sites to Cloudflare Pages and connected ***, so every push triggers a build and deployment without needing a separate GitHub Actions workflow. For some sites, SSR routes failed inside workerd and I ran into issues with legacy Pages Router/Emotion dependencies, and Node runtime and CSS bundling so I moved those apps to AWS Amplify. Learning: Cloudflare Pages & workers is lower cost, but its opennext adapter might not handle everything. Make sure to test SSR and auth before moving.

• 
  Tushar Mehta (@tushaarmehtaa) reported 54 minutes ago

  Openrouter is easily one of my favourite and most innovative non-foundation-lab AI companies. and when it got acquired for $8B, it felt like a very well-deserved outcome. but you might still think: $8b for a model router? one API → every model useful sure but increasingly commoditised vercel has AI gateway. cloudflare has AI gateway. we have a bunch of Indian versions of the same idea too. but before putting @OpenRouter in the same bucket as every other AI gateway, look at what they actually built for the ecosystem: > auto exacto: continuously measure which provider is actually running the same model best, then move traffic there automatically > fusion: turn disagreement between multiple frontier models into a signal, then use a judge to find consensus, contradictions and blind spots before producing the final answer. > advisor + subagent: let models dynamically escalate hard work to smarter models or delegate easy work to cheaper ones and there’s a bunch more i haven’t even included. openrouter started with “one API for every model” what they built on top of that is much more interesting.

• 
  Mixin (@MixinMessenger) reported 56 minutes ago

  The wider pattern: the stack is splitting into systems that accumulate identity and systems built to hold less of it. Signal shipped Automatic Key Verification on Aug 11, with Cloudflare and Trail of Bits as independent auditors. Vitalik argued on Aug 7 that pseudonymity no longer survives AI-driven metadata correlation, and that message-level unlinkability is the standard worth chasing. Different projects sit at different points on that curve — the direction of travel is what matters: collect less, retain less, expose less.

• 
  CapyToolkit (@CapyToolkit) reported 56 minutes ago

  @mdanassaif @spaceship Jump to Cloudflare! At least we all will be on the same boat when it's down... 🤪

• 
  Mike Choi (@guard_if) reported 59 minutes ago

  if you want to run a series A startup: Rails = the entire app. Postgres = source of truth. ClickHouse = analytics. Kafka = getting one into the other. (MSK) Sidekiq = everything async. ECS + EKS = deploys your own models. (GPUs, lol) Terraform = all of it, in code. Grafana + Sentry = knowing what broke. Statsig = flags + experiments. Stripe = money in. Mercury = money out. Loops = email. Intercom = support. Cloudflare = DNS. Vercel = marketing website. Linear + Graphite = shipping. Claude = coding / artifact generation total monthly cost: definitely not $20.

• 
  Jorge Trujillo (@zo0r) reported 1 hour ago

  @ritakozlov @Cloudflare @tan_stack it seems they tripled the amount of LOC, how is that better? from the outside, it looks more like poor architectural decisions beforehand that were worsened by the migration. title is also misleading, should be "next.js to tanstack migration"

• 
  Rock (@Rock4754) reported 1 hour ago

  @retiredbhondu Go to ur network settings and in DNS-> Private DNS hostname add -> 1dot1dot1dot1.cloudflare-dns. com *Remove the space before .com when entering it