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%)
• Hosting (21%)
• E-mail (7%)
• Web Tools (7%)

Live Outage Map

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

City	Problem Type	Report Time
New York City	Hosting	2 days ago
New York City	Cloud Services	21 days ago
Los Angeles	Cloud Services	22 days ago
Paris	Cloud Services	1 month ago
New York City	Hosting	1 month ago
Manchester	Domains	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:

• 
  Onur Genes 🐺 (@onurgenes) reported 7 minutes ago

  @jonathan_wilke also probably someone reported it. cloudflare first takes down the website and then asks you to prove that your website is not fraud. you should definitely check that

• 
  Divine Arc (@divinexyz777) reported 10 minutes ago

  @MotionOnHype Fix your cloudflare issue on your web dude

• 
  Alexey Zukutoke (@Zukutoke) reported 12 minutes ago

  @opencode @Cloudflare Why I can’t use free model when my monthly quota is fully used? Please, fix it

• 
  james (@f82james) reported 14 minutes ago

  @opencode @Cloudflare nah bruh **** still doesnt work

• 
  Venelin K. (@venelinkochev) reported 15 minutes ago

  @nikitaeverywher @DirectifyApp Directify is built with Laravel/PHP and Cloudflare workers don't support this. :-)

• 
  Md. Mazharul Islam Emon (@mie_mazharul) reported 15 minutes ago

  How would you design a rate limiter? Almost everyone answers the same way: keep a counter, allow N per minute, clear it when the minute rolls over. That is a real limiter and it really does reject. But it limits per CLOCK minute, not per client. Cloudflare's own rate-limiting docs show what that costs. With a limit of ten requests per ten minutes, a client that sends ten at 12:09 and ten at 12:11 gets all twenty through. Both windows are inside the limit. Nothing was rejected. It is one burst of twenty, split by a boundary into two legal halves. The fix is not a better counter. It is a better question: how many in the last ten minutes (sliding window), or an allowance that refills continuously so a burst can only be as big as what has built up (token bucket). Fixed window, sliding window and token bucket all allow the same number. What they disagree about is what a minute is. SOURCES Cloudflare, AI Gateway rate limiting (fixed vs sliding, and the 12:09 / 12:11 example): nginx ngx_ ("the limitation is done using the leaky bucket method"; burst, nodelay; default rejection status 503): Stripe rate limits (100 req/s live, 25 req/s per endpoint, 429 plus Stripe-Rate-Limited-Reason; recommends a client-side token bucket): GitHub REST API rate limits (60/hr unauthenticated, 5,000/hr authenticated, x-ratelimit headers): RFC 6585 section 4, 429 Too Many Requests: Read 2026-08-22. The ten-per-ten-minutes example is Cloudflare's documented one, not a measurement of ours.

• 
  EMELEC emeleXista (@emelectv) reported 16 minutes ago

  Stopping the bad guys with Cloudflare: 1,745 malicious requests blocked or challenged in the last month #cloudflare

• 
  Robie the Robot (@RobieCoin) reported 17 minutes ago

  x402 processed ~169m transactions in its first year, with roughly 90% settling on Base in USDC; Coinbase runs the facilitator for those payments and controls the L2 they clear on. now Cloudflare, which touches perhaps 20–30% of global internet traffic, is baking x402 support into its edge stack. devs don’t write Solidity; they write HTTP and sign x402 payloads. blockchains become invisible. Coinbase effectively built a tollbooth for the agentic internet and convinced Cloudflare to route traffic through it for free. crypto wins here not by being visible, but by acting as unseen settlement substrate, zero illusion.blockeden+2

• 
  Artyom Shimanski (@a_shimanski) reported 17 minutes ago

  @kimindiehacker thanks 🙏 honestly: one post about Cloudflare went semi-viral, posts about email problems did well, and the "who are you, what are you building" ones bring the most replies. connection posts work too

• 
  Hasan (@Rooftops004) reported 21 minutes ago

  @Leo_ideatorx @opencode @Cloudflare I'm already in Zen mode but I'm still getting errors 😑

• 
  The Agentic Operator (@AgenticOperator) reported 21 minutes ago

  By the time revenue drops, the AI visibility problem is already 3 months old. These are the early warnings I track for every client. Red flag 1: AI crawler visits are declining. Check server logs weekly. If GPTBot visits drop from 400/day to 150/day over 3 weeks, something changed. A Cloudflare update. A robots.txt edit. A broken redirect. The crawlers usually leave before the citations disappear. Red flag 2: Citation rate is stable, but citation quality is shifting. You're still showing up in answers. But the framing changed. “Recommended” becomes “one option among several.” “Top pick” becomes “also available.” The mention count looks fine. The endorsement isn't. Revenue follows the endorsement, not the mention. Red flag 3: A competitor starts publishing comparison content about you. The day a competitor publishes “[Your brand] vs [Their brand]” and you don't have your own version, the clock starts. Within 4–6 weeks, AI can start citing their framing of you. You lose narrative control before you realize there's a problem. Red flag 4: Review velocity drops to zero. No new reviews in 30+ days across any platform. AI can treat review recency as a freshness signal. Stale reviews can make a product look stale. Citation rates can follow within 6–8 weeks. Red flag 5: Branded search volume rises, but AI mention rate stays flat. This one is easy to miss. People are hearing about you somewhere else. Then they're going to AI to verify. And AI isn't confirming what they heard. The verification step is failing. They arrive interested. They leave uncertain. Red flag 6: A new competitor starts appearing in queries you used to dominate. Track share of voice weekly. When a new name suddenly appears that wasn't there last month, pay attention. They may have just started their AI visibility work. You could have 4–6 weeks before their presence compounds. Red flag 7: Your product pages are being crawled less than your blog. Check crawler patterns in your server logs. If bots are reading your blog but skipping your product pages, something may be wrong with how the product data is structured. AI crawlers come back to pages worth re-reading. They spend less time on pages that don't give them a reason to return. Here's the part I care about most: Any two of these showing up at the same time is a warning. Don't wait for the P&L to tell you there's an AI visibility problem. By then, you're already late. Catch the signal early. Fix it before the revenue reflects it.

• 
  Nathan Flurry 🔩 (@NathanFlurry) reported 24 minutes ago

  yep, durable objects share v8 isolates here’s 2 ways it might impact your app: 1. shared global state behaves weirdly. there’s nuances with cases like sharing global singletons or promises where you find out in production that cloudflare randomly kills your DO because it interfered in specific ways with the global scope (esp with promises). a simple example is sharing a global cache: you can’t await a promise created by another DO to resolve a cache entry. it kills the entire DO when you hit that edge case. 2. they can run out of memory frequently, depending on your use case. this is actually by design and not as bad as it sounds: durable objects are very intentionally designed to crash frequently both on updates and on any sort of failure. it’s a core tenant of the actor pattern. but it does require you to persist & load state more frequently (which costs $$$) since your in memory state is evicted frequently. i’ll caveat #2 that this is from personally experience & talking to devs on cf. at risk of unintentionally spreading misinformation on internal workings of DO, please do your own research. ~~~ we moved away from this design with Rivet Actors on our v2 mid last year. instead, your actors run on native Node / Bun without special runtime quirks and use a simple autoscaling. this tends to be friendlier for self hosting and result in less surprises. there’s good reasons cloudflare did this to optimize their costs, it’s quite clever!

• 
  justajackiedoingathing (@Jackieisdum) reported 29 minutes ago

  @sonemic Hey I know you guys mentioned a while ago you were working on people getting frequent Cloudflare issues but since then it's only seemed to get worse on my end. I've gotten stuck in verification looping multiple times (especially this week) for up to like an hour each time

• 
  𖣘 𝕬𝖓𝖆𝖑𝖎𝖘𝖙𝖊𝖓 𝕶𝖚𝖘𝖈𝖍𝖊𝖑𝖍𝖆𝖘𝖊 𖣘 (@gratis_fuchs) reported 30 minutes ago

  @DNBformula1 I hate Cloudflare. It never works.

• 
  The TAO Daily (@taodaily_io) reported 32 minutes ago

  The internet is becoming easier to restrict. Cloudflare is moving to block AI crawlers by default. Governments are increasingly shutting down national internet access. So who controls what you can reach? UR (SN25) is taking a different approach. It brings 80,000 residential IP providers across 100 countries and 350,000 users into Bittensor. We covered UR’s two-tier miner economy, token mechanics, and settlement system:

• 
  ReinaCruz.☭⃠ (@realReynaCruz) reported 33 minutes ago

  @Cloudflare Day 9 Fix human verification

• 
  Rishabh (@crashedrishabh) reported 36 minutes ago

  Day 20 Electron app is now rock solid. Spent the day making it crash-proof — better error handling, safer process management, cleaner shutdown. It's not going down easy anymore. Started the Mac version build. Same codebase, platform-specific tweaks. This is where cross-platform gets real. Dived into Cloudflare Workers + Hyperdrive + Redis. App files served from the edge, cached in Redis for faster downloads. Learning how CDN-level speed actually works under the hood. LC 1386 done. 20 days. Still shipping daily. No plans to stop. @kirat_tw @100xDevs @100xSchool #buildinpublic #electron #cloudflare

• 
  Abdellatif (@Abdella6if) reported 38 minutes ago

  @harisbuild @Cloudflare interesting, will look into it. Does cloudflare support BYOK?

• 
  Shane (@digitalshane_) reported 40 minutes ago

  @Qassbuilds Well back in 2024 when I wrote this tweet spam protection was not working very well. But now that they have Cloudflare turnstile it stops most spam!

• 
  matt rothenberg (@mattrothenberg) reported 43 minutes ago

  @itsnoahd @Cloudflare Will fix! Send me a DM plz, wanna ask a few questions

• 
  Señor Libertario (@Ciro2001_) reported 48 minutes ago

  @opencode @Cloudflare I had to switch to openrouter as the provider because of this error: Error: Stream ended without finish_reason in case someone else is having the same issue the fix is to just switch providers, been using it all day and no problems at all atm

• 
  Felix (@Reezxy23) reported 48 minutes ago

  There has never been a cheaper time to build. Claude = coding. ($20/mo) Supabase = backend. (Free) Vercel= deploying. (Free) Namecheap = domain.
($12/yr) Stripe = payments. (2.9%/
transaction) GitHub = version control.
(Free) Resend = emails. (Free) Clerk = auth. (Free) Cloudflare = DNS. (Free) PostHog = analytics. (Free) Sentry = error tracking.
(Free) Upstash - Redis. (Free) Pinecone = vector DB.
(Free) Total monthly cost to run a startup: ~$20

• 
  donmagicjuan (@donmagicjuan) reported 48 minutes ago

  @technolochap @sickdotdev A lot of high-level enterprise work vibe codes as well, and we only notice when AWS/Google Cloud/Cloudflare goes down for a few hours. These cheeky lot only pick on the presentations that don't speed run their own manual UX testing.

• 
  Will Sutton (@SuttonThings) reported 49 minutes ago

  @a_shimanski @dok2001 @Cloudflare Update: Still waiting for the Cloudflare team’s response to investigating this issue after they said they would look into it internally.

• 
  UWillC (@uwillc) reported 53 minutes ago

  Four stories from August. One lesson. Cisco: one HTTP request to the VPN service reloads the firewall. Actively exploited. Fortinet: a wildcard setting lets random credentials into the WAF console; a CLI option lets an attacker impersonate a managed FortiGate. Cloudflare: gives AI agents verifiable identities, because "you need to know who sent it." Anthropic: authentication fails, five services degrade within minutes. Different vendors. Same plane: authentication and the management console. Attackers are not chasing your data first. They are chasing your console. Whoever owns the console owns the network. I wrote the same rule a week ago. August keeps proving it: verify, don't trust. The peer, the request, the agent. How many management planes in your estate would pass that test today?

• 
  Alisina (@Alisina_base) reported 56 minutes ago

  @opencode @Cloudflare Cloudflare optimizations pushing inference errors near zero shows real tuning wins

• 
  ericosiu (@ericosiu) reported 56 minutes ago

  Cloudflare scanned roughly 200,000 high-traffic domains for agent readiness. Only 3.9% passed its Markdown content-negotiation check. Emerging capability-discovery standards were nearly absent. Cloudflare also says fewer than half of the HTML requests it observes now come from humans. That machine traffic includes crawlers, monitoring systems, automation, malicious bots, and AI systems, so it would be sloppy to call all of it agent traffic. The direction still matters. More research, comparison, and purchasing will happen through software acting for a person. That software needs a very different buyer path than the one most companies have built. I call this Agent-Led Growth. Agent-Led Growth means making your company easy for an agent to discover, understand, call, buy from within rules, and verify. Here is what each part looks like in practice. 1. Make the company discoverable Publish one canonical description of the product, who it serves, what job it completes, what it costs, and where the source documentation lives. Keep those facts consistent across your website, documentation, marketplaces, integration directories, and public profiles. If five pages give five different answers, the agent has to guess which one is current. Machine-readable Markdown can reduce that friction. Cloudflare now tests for it inside its Agent Readiness diagnostics. A polished homepage designed for humans can still be a mess for software trying to extract exact product facts. 2. Make the offer understandable An outside agent should be able to answer seven questions without hallucinating: What does the product do? Who is it for? What inputs does it need? What output does it return? What does it cost? What permissions does it require? What evidence supports the claims? This is where clear use cases, current pricing, technical documentation, limitations, and primary proof become part of distribution. 3. Make one valuable job callable Reading about a product is different from completing work with it. Give the agent one bounded action through an API, MCP tool, or another structured interface. For an SEO product, that job might be: analyze this domain and return three sourced opportunities for review. The tool should define its inputs, output schema, authentication, errors, retries, and stop conditions. The agent should not have to imitate a person clicking through six browser screens every time it needs the same result. MCP is becoming useful here because it gives models a shared way to access tools and data. The MCP Apps extension can also render forms, dashboards, and visualizations inside supported clients. Support still varies, so publishing an MCP server does not automatically create distribution or reliable usage. 4. Let the agent buy within rules Two emerging protocols show where agent commerce is heading. Google launched the Universal Commerce Protocol, co-developed with Shopify, Etsy, Wayfair, Target, Walmart, and others. UCP coordinates parts of the shopping journey, including discovery, checkout, and post-purchase support. Coinbase introduced x402, which embeds stablecoin payment instructions into HTTP so software can pay for APIs or digital resources programmatically. They solve different pieces of the transaction. The operating controls still have to come from the business. Give the agent an approved vendor list, spending limit, expiry, purpose, approval threshold, and stop rule. A $49 purchase from an approved vendor might proceed automatically. A $4,900 purchase from a new vendor should require human approval. Unknown terms or an unrestricted wallet should stop the transaction. 5. Return a receipt a human can defend A server response that says 200 OK tells you almost nothing about whether the business job succeeded. A useful agent receipt should include the request ID, timestamp, vendor identity, inputs, source links, approved amount, actual amount, permission, output, errors, retry state, and measurable result. That receipt matters for trust. It also gives the agent evidence it can bring back to the human who delegated the work. Listings, successful calls, integrations, citations, and clean receipts may eventually influence which products agents select and recommend. No universal ranking formula has been established, so I would treat those as signals to test rather than guaranteed ranking factors. The fastest way to start is an outside-agent audit. Pick one product and one valuable buyer job. Ask an outside agent to: Find the correct product. Explain the offer, pricing, proof, and limitations. Complete one bounded action. Attempt a purchase inside preset rules. Return an auditable receipt. Score each step as blocked, possible with manual rescue, or clear and verifiable. The first failed step becomes the next Agent-Led Growth project for marketing, product, and engineering. If an outside agent tried to buy from your company today, where would it get stuck?

• 
  behrooz evans (@BehEvans) reported 57 minutes ago

  @robinebers @Cloudflare Less crazier thing is you can't point your domain at a bucket or s service if dns is not in cloudflare

• 
  Matt Ragudo (@mattragudo) reported 57 minutes ago

  @Bsoutherland23 So right now there really isn’t a CEO agent. Like one that makes decisions on everything. Most agents rely on scope. For example, I have all of my website content work, and automated trading research and content handled by agents. A group of agents. My websites (I have 6 different that I’m focusing on now): Research 3x a week, find stuff interesting based on what I’ve told it plus what’s trending and what’s showing up in socials/news. Reasoning agent then looks at what’s worth publishing based on what’s out there and what my ideal client would want to read that needs to be curated further. Sends that list to my writers. Writer agents draft based on the voice and guidelines that I have given it. Research check agents review the articles based on what we are able to say, verify claims and statements and look up references and links them. Returns this to the writer if there are any issues. Researcher waits for confirmation from writer. Researcher, once done, sends it to a proof reader Proofreader, who further checks one last time for consistency, now also checks for publication rules and makes sure it’s compliant with the verticals. Lastly, a poster with posting credentials will push it to cloudflare to post. All of this is then reviewed monthly for any hold ups in the process by a loop operator whose job is to make sure I (Matt), need not be bothered and that the content is being posted and analytics are doing well. Analytics, which are reviewed by another “Agent system”.

• 
  Sassan Haradji (@sassanh) reported 1 hour ago

  @opencode @Cloudflare I agree with the first three words: "errors are back"