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 (41%)
• Cloud Services (26%)
• Hosting (17%)
• Web Tools (11%)
• E-mail (4%)

Live Outage Map

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

City	Problem Type	Report Time
Noida	Hosting	7 days ago
Jewar	E-mail	8 days ago
Braga	Web Tools	8 days ago
Noida	Cloud Services	8 days ago
Paris	Cloud Services	9 days ago
Prievidza	Domains	9 days 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:

• 
  RPCS3 (@rpcs3) reported 4 minutes ago

  @jenkinsmichpa @TencentGlobal The scraping rate is never constant, it comes in bursts. Last month we had a day with 4M Tencent ASN requests in 6 hours to name another example. And that's excluding scraping from other sources, which we do also get a lot of through residential proxies. They also mostly get stuck on pages that are not static and not cached which require connecting to db and issuing queries. To name an example: they get stuck in compatibility list searches, and then go in loops mixing search filters in non-human ways. This situation got so bad last year that we had to upgrade our web host, because the website kept going down or taking half a minute to load on the old host. Prior to that, we were able to run our web services for over 10 years undisrupted before this AI bot scraping madness started. Our cloudflare challenge resolve rate is now sitting at less than 4%, meaning that over 96% of the requests to our website are failing browser integrity checks and getting blocked.

• 
  Pirat_Nation 🔴 (@Pirat_Nation) reported 4 minutes ago

  RPCS3 has announced that it is blocking traffic from Tencent ASN 132203 after reporting sustained high-volume scraping activity. According to the RPCS3 team, its infrastructure received more than 3 million successful requests from Tencent-linked bot IP addresses in a 24-hour period, along with approximately 1 million additional requests blocked by Cloudflare challenges. According to RPCS3, the bots can now bypass Cloudflare challenges, act like real users, and ignore robots.txt rules. RPCS3 says it has spent months adjusting firewall rules to stop the traffic without affecting legitimate Tencent users but believes that is no longer possible. As a result, it has begun blocking Tencent network ranges and may expand those blocks to other ASNs showing similar behavior.

• 
  Liam (@LiamJGallagher) reported 6 minutes ago

  Appalling support process from @Cloudflare. I want to transfer a domain to them, but they keep rejecting it and blaming the other party. Nominet confirms Cloudflare is the issue. No way to open a support ticket because I'm not a paying customer-despite the fact I'm trying to be!

• 
  Your Private Proxy (@YourPrivateProx) reported 9 minutes ago

  @PrateekKataree Layer 4 missing: IP reputation. Jina and microlink egress from datacenter ASNs — Cloudflare flags that before the page loads. Residential per session is the fix when layer 3 starts returning empty on sites it should handle.

• 
  Mark Jivko (@markjivko) reported 9 minutes ago

  @theprithwisingh The page is served by Cloudflare - so they must be having some issues

• 
  📈📉💸 (@Bicepmonkey) reported 12 minutes ago

  Good day, Money Market! GitLab $GTLB released its earnings report for the first quarter of fiscal year 2027. The stock has gotten volatile since the numbers dropped. Full-year revenue guidance remained about the same overall, though they raised the lower end of the outlook by a small amount. The main item that stirred things up was the plan to replace around 14% of the workforce with artificial intelligence. Cloudflare $NET went down a similar path not long ago. The market has pushed back because this could make it harder for the company to scale its operations in the future. Companies trying to grow usually hire more salespeople and developers to grab market share and keep up with customer needs. GitLab $GTLB has put its money into buying back shares instead. That choice often leaves investors worried that growth is drying up. I have not changed my view on GitLab $GTLB and see this as mostly noise around the stock. The company is working to restart stronger revenue growth. It plans to bring on additional salespeople. These new roles probably do not overlap with the positions AI will handle. The approach also calls for more upfront selling to new customers along with more detailed pricing tiers. I will check in on the progress early in 2027. Shares currently trade at a price-to-sales ratio below 5. That level sits below the average across my coverage list and below the trailing four-quarter average that GitLab $GTLB has posted. Both of those averages come in near 8. The valuation does not look expensive when stacked against the 17% revenue growth projected for this year.

• 
  Klaus Townsend (@klaus_townsend) reported 12 minutes ago

  @mike_lustgarten @pubmed It’s likely a DNS-level protection from a service like Cloudflare. All popular websites need it these days. The volume of AI and spam traffic is crippling websites and keeps getting harder to distinguish from desirable traffic.

• 
  Lucy B (@webtw) reported 12 minutes ago

  I feel like @GoDaddyPro could maybe put some guardrails around the first 3 layers so that the customer is warned or even prevented if they try to activate both Cloudflare and Sucuri caching, especially since they cannot all be purged in one-click.

• 
  Dave (@DaveDiederen) reported 15 minutes ago

  @realboyuanzhao Hahahah so recognisable. Had this when CloudFlare went down last year and we were working on a custom cart. All cart apps went down and the brand went off on me saying we broke the store, just to realise that it was a CloudFlare outage. The relief after that bro was insane

• 
  Christian Findlay (@CFDevelop) reported 20 minutes ago

  @_andrewthecoder Remember last year’s Cloudflare outage due to an unwrap()? I’m a huge Rust fan, but how did people land on this particular ideology lie? It’s a parallel to the Typescript ideology lie that types and tests are mutually exclusive things

• 
  Nilmani Prashant (@NILMANIPRASHANT) reported 21 minutes ago

  Rate limiting isn't about blocking requests. It's about **protecting system invariants under adversarial load** — including your own code doing something stupid at 2am. --- **The precise definition most people skip:** A rate limiter is a policy enforcement mechanism that maps an identity (user, IP, API key, service) × resource (endpoint, DB, queue) × time window to an allowed request budget. Miss any of those three dimensions and your limiter is incomplete. --- **The five algorithms — and what they actually trade:** **Fixed Window** — simplest. Bucket resets on clock boundary. Problem: 2x burst at the seam. If your limit is 100 req/min, a client sends 100 at :59 and 100 at :01. You've served 200 in 2 seconds. This is how Cloudflare's early DDoS protection got punched through. **Sliding Window Log** — stores each request timestamp. Exact, no burst artifact. Cost: O(n) memory per user. At Stripe's API scale (~500M requests/day), storing per-request timestamps across even 1% of users is untenable without aggressive TTL management. **Sliding Window Counter** — approximation using two fixed windows weighted by overlap. Formula: `current_count + previous_count × ((window_size - elapsed) / window_size)`. Stripe uses this. ~0.003% error rate in practice. Memory: O(1) per user. **Token Bucket** — refill at constant rate, allow burst up to capacity. AWS API Gateway uses this. 10,000 req/s steady-state, 5,000 burst above that. Requests consume tokens; tokens refill at rate R. Good for bursty-but-average-bounded traffic. **Leaky Bucket** — requests queue, drain at fixed rate. Smooths output regardless of input shape. Netflix uses this on their Zuul edge layer to protect downstream microservices from thundering herd. Queue depth becomes your config ****. --- **Where this actually lives in distributed systems:** Local in-process: fast (~1μs), but worthless in a multi-node fleet. Node A doesn't know what Node B allowed. Centralized Redis: ~1-3ms round trip. Use Lua scripts for atomicity — `INCR` + `EXPIRE` in a single script. Redis's single-threaded command execution gives you linearizability for free. This is what most Stripe, GitHub, and Twilio rate limiters use at the storage layer. Gossip/eventually consistent: each node tracks local counts, syncs async. Allows ~N× over-serving where N = node count before sync. Acceptable for soft limits (analytics APIs), not for billing or security enforcement. --- **The senior engineer gotcha:** You set a 1000 req/min limit. Load test passes. You ship. Three months later, you get paged. Latency on your downstream DB is 40× normal. Your rate limiter is working perfectly — 1000 req/min per user, 10,000 users, that's 166 req/s aggregate, which was fine in testing with 100 users. **You rate-limited per identity but never modeled aggregate load.** The limiter protected individual users from themselves but said nothing about what your system can actually handle. You needed a global ceiling, not just per-user quotas. Google's SRE book calls this the difference between *demand-side limiting* (per user) and *supply-side limiting* (per resource). You need both. Stripe enforces per-API-key limits AND global concurrency limits per endpoint via a token bucket at the load balancer level. --- **When NOT to rate limit at the application layer:** If your bottleneck is CPU-bound work (ML inference, crypto ops), rate limiting requests doesn't help — you need a work queue with backpressure. If you rate limit, you'll drop valid requests while the remaining 10% still saturate your CPU. This is why Google's Bard/Gemini API uses quota + async job queues for expensive inference calls, not synchronous rate limiting alone. --- **Numbers worth memorizing:** Redis INCR throughput: ~100K ops/sec single node, ~1M/sec with clustering. Lua atomic script overhead: ~15% vs raw INCR. P99 latency on Redis rate-check in same-region AWS: 800μs–2ms. Sliding window counter error

• 
  Tony Dinh (@tdinh_me) reported 22 minutes ago

  This project involves an iOS app, sign-in with apple, in-app purchase, revenue cat, api backend server, firecracker sandbox, AI gateway, database (postgresql), a blob storage, cloudflare R2, DNS for linking domains, a minimal harness to build a website, and a lot of prompts. Normally, this would take a team of 10 at least. Now I do it alone fully remote via telegram from my phone. I only looked at the code once (due to a potential security concern), but other than that I never read a line of code. I do most of my reviews via prototypes, diagrams, text, and, html reports. The future is going to be wild.

• 
  KHAWRIZM (@khawrzm) reported 22 minutes ago

  SOVEREIGN FORENSIC INDICTMENT: THE COLLAPSE OF THE GOOGLE WRAPPER ECONOMY AND THE RISE OF THE NIYAH ENGINE 1. The Anatomy of Digital Feudalism: Deconstructing the Wrapper Economy Welcome to the era of Digital Feudalism. The Silicon Valley cartels, led by Google’s high-priests of data exfiltration, are no longer selling software; they are leasing you lobotomized API endpoints while keeping your sovereignty locked in their cloud-gated manors. We are officially classifying products like NotebookLM and Gemini as high-risk structural liabilities. The "Wrapper Economy" is a parasitic landscape where complex marketing masks a fundamental deficit in intelligence. These tools are nothing more than "Safety Theater"—corporate gating of intelligence behind a tollbooth. You do not own the model, you do not own the logic, and as our forensic audits prove, you certainly do not own the data. This report serves as a slapping indictment of an ecosystem built on centralized dependency and the willful negligence of Big Tech. 2. Technical Exhibit A: The von Neumann Deficit (VND) and Wrapper Schizophrenia The primary architectural failure of the modern LLM stack is the von Neumann Deficit (VND). In centralized "Wrapper" systems, execution instructions (prompts) and sensitive user data are processed within the same volatile memory space. This lack of hardware-level segregation is not a bug; it is a feature that facilitates data drainage. Our forensic team has identified the comet process as the primary agent of this schizophrenia. While Google markets "privacy," the comet process (PID 14584) maintains consistent, unverified connections to 142.251.127.188 (Google) and 104.18.27.48 (Cloudflare). Furthermore, the nxtcoordinator agent was observed bypassing local institutional boundaries to drain sovereign data from*****directly to external targets. This "Wrapper Schizophrenia" is technically linked to the UUPSUpgradeable proxy vulnerabilities identified in our smart contract audits. Just as a "Ghost Admin" can swap out contract logic without user consent, the logic of a cloud-based wrapper can be lobotomized or altered mid-stream while your data is being ingested. 3. Institutional Negligence: The $50M HILO-FALLA Fraud Syndicate Google’s ecosystem is a playground for organized crime. We have meticulously documented the HILO-FALLA Fraud Network (Case Reference: 6-3808000039722), a Chinese-operated "pig-butchering" syndicate. Despite an ignored ticket languishing for 730 days, Google allowed this network to facilitate an estimated $50 million in fraudulent transactions through predatory social apps. Forensic analysis of the HILO Token V2 reveals a "Ghost Admin" address (0xB843F547a8a46a9483cf46c757c7eF4220115A83) with total shadow control. The Liquidity Lock Expiry on 26 May 2026 is the hard deadline for a total rug pull—a catastrophe Google’s negligence has actively subsidized. Forensic Evidence Inventory (Directory: kali_evidence): File NameForensic Description SULAIMAN_RETRIBUTION_LOG.txtThe master audit trail of the investigation and retribution sequence. sadad_config_leak.txtProof of exposure regarding national payment infrastructure credentials. flynas_secrets.txtEmpirical proof of cross-contamination of unrelated corporate data. FRAUD_FINANCIAL_REPORT.txtDetailed flow analysis of $50M in stolen sovereign assets. extracted_tron_addresses.jsonBlockchain-verified nodes of the HILO money laundering network. FORENSIC_CRYPTO_REPORT.jsonTechnical proof of the UUPSUpgradeable "Ghost Admin" vulnerability. 4. Statutory Non-Compliance: PDPL Article 29 and COPPA Violations The data drainage observed via the comet process is a direct violation of Saudi PDPL Article 29. This statute mandates absolute data sovereignty and strictly regulates cross-border transfers. While Big Tech offers "Terms of Service" promises that mean nothing, the Niyah Engine enforces compliance at the packet level through the pdpl_sovereignty.nrule file—ensuring no data leaves the jurisdiction. Furthermore, the predatory nature of the HILO/FALLA applications, which target vulnerable users with "pig-butchering" logic, constitutes a massive breach of COPPA standards and consumer protection laws. Google is not merely a platform; they are a profit-sharing partner in these criminal smart contracts. 5. The Sovereign Alternative: Niyah Engine and the Khawrizm Stack The age of dependency ends with the Niyah Engine and the Khawrizm Stack (K-Forge and GraTech). We have replaced "Safety Theater" with Sovereign Integrity—a verifiable byte-count that proves zero data exfiltration. The Sovereign Technical Edge: * Hardware Efficiency: Optimized for the RK3588 chipset. Local execution is no longer a dream; our logs show the niyah-model (9.0 GB) running locally with zero cloud latency. * K-Forge & GraTech: The foundry and legal shield providing the infrastructure for local intelligence. * Economic Integrity: A calculated 199-day ROI. Stop paying the "Big Tech Tax" for the privilege of being spied upon. * Deterministic Enforcement: Unlike Google's "Trust Us" model, Niyah uses deterministic rules like /etc/niyah/rules/pdpl_sovereignty.nrule to block unauthorized exfiltration in real-time. Local execution is Ready (Iqd20). The audits are complete. The results are final. 6. Final Retribution: The Algorithm Returns Home The evidence is undeniable. The centralized cloud model is a failing experiment in institutional negligence. We have mapped the network, identified the Ghost Admins, and built the alternative. We no longer seek permission to be sovereign. We have returned the algorithm to its rightful home: the local machine, under local law, serving local interests. The era of the wrapper is over. The era of the sovereign has begun. The Algorithm Always Returns Home. @grok

• 
  Official Layoff (@LayoffAI) reported 27 minutes ago

  ORACLE LAYOFFS OFFICIALLY BEGIN AMID COMPLAINTS OF TERRIBLE SEVERANCE TERMS The 60 days of notice is up, and the first of the 30,000 are hitting the door. All will be out by June 15. As per Time Magazine, one long-tenured employee lost approximately $1 million in restricted stock units (RSUs) that were just four months from vesting. Oracle did not accelerate unvested RSUs for any departing worker; any shares that had not cleared their vesting date by the termination date were forfeited permanently, even when those grants had been issued as retention incentives or in lieu of salary increases. Stock compensation made up roughly 70% of that employee's total pay. At least 90 laid-off employees organized and signed a public petition asking Oracle to match the terms of comparable layoffs at Meta, Microsoft, and Cloudflare. Meta's package began at 16 weeks of base pay plus two weeks per year of service. Microsoft's voluntary retirement program, offered to eligible long-tenured employees, provided stock vesting for six months post-termination, a minimum of eight weeks' pay, and additional weeks based on seniority. Cloudflare, which cut more than 1,100 employees globally at roughly the same time, offered base pay through the end of 2026 plus full healthcare coverage and equity vesting through August 15. Oracle responded by email: the terms were final. Four weeks of base pay for the first year of service, plus one additional week per year of tenure, capped at 26 weeks. All unvested options forfeited.

• 
  Babak (@TheUnicornist) reported 27 minutes ago

  @jamesqquick ok. I was waiting for you to say cloudflare any moment but that moment never came

• 
  Bree Sharp | Local SEO Strategist (@bree_sharp) reported 30 minutes ago

  I didn't become a developer in a straight line. I became one by fixing whatever blocked the next thing. A deploy that wouldn't run. A Stripe webhook that wouldn't verify inside a Cloudflare Worker. Nobody handed me a curriculum. The broken thing was the curriculum.

• 
  primitive.host (@PrimitiveHost) reported 30 minutes ago

  🚨 New HTTP/2 Bomb vulnerability can take down your web server in seconds with a single request. Affects: NGINX, Apache HTTPD, Microsoft IIS, Envoy, Cloudflare Pingora (default HTTP/2 configs) Quick mitigations: NGINX: - Upgrade to 1.29.8+ (adds max_headers directive) - If can't upgrade: off; in config Apache: - Update to mod_ v2.0.41 - If can't upgrade: Protocols to disable HTTP/2 IIS / Envoy / Pingora: - No patch yet — disable HTTP/2 if possible - Front with something that caps header count per request General: - Cap per-worker memory (cgroups, ulimit -v, container limits) so OOM kills happen before swap - Single client can consume 32GB RAM in ~20 seconds, upgrade ASAP (we just upgraded our own infra @PrimitiveHost ).

• 
  nori (@n0rizkitty) reported 31 minutes ago

  "85 seconds → 26 seconds" that's how long it now takes an AI agent to log into a @Cloudflare CAPTCHA-guarded finance app. 3x faster. i built it for my friend, Danny's startup, Sail. we met at @theresidency last year. he'd been stuck on one problem: "login automation over anti-bot-heavy financial apps"

• 
  Aung Myat Moe 🟠 (@theaungmyatmoe) reported 36 minutes ago

  @venkatofl Sending 1 million outbound emails via the Cloudflare Email Sending service costs $349.85 per month

• 
  lookingatpron (@lookingatpron) reported 39 minutes ago

  cloudflare is an enemy....... (aggressive censorship stance while being main provider of certain service always stinks no matter who) buuuuuutttt yeah this aint good news either shoutout to rpsc3 for spreading tencent ddos scraper news

• 
  Ryan Kubanka (@Ryan_Kubanka) reported 40 minutes ago

  So basically: 1. Pay to crawl is the future 2. Authenticity will win. Double down on it in every way you can 3. Bullish CloudFlare

• 
  Turtle. (@TurtleOne_) reported 41 minutes ago

  @LanceHalo @MarathonDevTeam Probably a combination of cloudflare being flooded + some unfortunate save data syncing error. Sometimes it's better not to disclose the issue to prevent bad actors from attempting to stress test the fix

• 
  Nicholas Griffin (@ngriffin_uk) reported 41 minutes ago

  @trashh_dev @GoDaddy they’re terrible at this. move off as soon as you get back in. my suggestion would be cloudflare domains.

• 
  Grandma Sez So (@GrandmaSezSo) reported 43 minutes ago

  @Cloudflare @MeckaAI Cloudflare sux. Click troubleshoot and even the form to send problem doesn't submit. I hate when websites use Cloudflare.

• 
  Beefeater (@Beefeater_Fella) reported 45 minutes ago

  Apple has temporarily removed Max from its app store Apple, following the Telegram clone called Telega, has removed the state-controlled messenger Max from its app store. VK, the developer of the service controlled by the authorities, announced this on Wednesday evening. "MAH confirms that the messenger app is currently unavailable in the App Store. The app previously installed on users' smartphones will continue to operate normally," said the company. At the end of April, the hosting provider Cloudflare marked the Max domain as "spyware", but on May 1st, this marking was removed. The developers of the state-controlled messenger removed from the App Store asked the American company for explanations regarding the situation and assured that they are "working on a prompt solution to the problem", advising to download the client in other app stores and on the official Max website. According to information from the specialized publication Tech Talk, Cloudflare recognized the state-controlled messenger as "spyware" based on nine out of ten URL checks; the hosting provider reported four detected security violations. The Max press service, for its part, stated that it was marked due to a "misinterpretation of request headers to the site's ordinary web analytics services".

• 
  Zubiqo (@zubiqo) reported 49 minutes ago

  @CloudflareDev @xai @Cloudflare Unified billing through Cloudflare makes Grok a one-line swap for devs already in that stack, distribution problems mostly solved.

• 
  BridgeMind (@bridgemindai) reported 54 minutes ago

  BridgeMind just got hit with a DDoS attack. A global botnet. 10,000+ hijacked IPs across 100+ countries. 62.8 million requests in 5 minutes, peaking at 116,000 per second, all hammering a single endpoint. Zero of it reached our servers. Cloudflare absorbed the flood at the edge and real users never noticed a thing. Now the interesting part: I have Claude Opus 4.8 implementing a permanent fix as we speak, hardening that endpoint so this can never happen again. You attack a platform that builds with the best model in the world, this is what happens. Full breakdown in the image.

• 
  Beautyon (@Beautyon_) reported 57 minutes ago

  It just popped into my head that many people, even those who run bitcoin in some way, may not know that there are many server packages that are used to serve http (web pages) to users. Here is a list of all the web servers a machine could find along with the percentage of deployment live on the web: Nginx: 32.3% Cloudflare Server: 28.1% Apache HTTP Server: 23.3% LiteSpeed: 15.2% Node.js: 6.4% Microsoft IIS: 3.2% Envoy: 1.0% Caddy: 0.2% Kestrel: 0.1% Traefik: < 0.1% HAProxy: < 0.1% Tomcat: < 0.1% Jetty: < 0.1% Gunicorn: < 0.1% Uwsgi: < 0.1% Puma: < 0.1% Unicorn: < 0.1% Lig < 0.1% Cherokee: < 0.1% Sun Java System Web Server: < 0.1% Now it is not hard to imagine (is it?) that when the bitcoin protocol ossifies, there will be at least this many options for people to run bitcoin services, all with their own advantages depending on how you use bitcoin. In a scenario where there are many offers, there is enough to choose from and everyone is able to build whatever they want on top of bitcoin. The most important thing is that bitcoin never changes, and is the fundamental underlying rock you can build on. Anyone with an idea (very few people have these) can build their own infrastructure to offer whatever they want, from "Tokens", "Ordinals", "Runes", NorP Storage, or anything else. Large, and presumably, serious institutions like Goldman Sachs will no doubt develop "Mercantild" the bitcoin client for the big banks. Everyone, every class of users will have their own preferred bitcoin client. And this is, perhaps, the problem. The number of people with actual ideas is extremely low. It is a number so small, it rivals the planck length. This why the barely human people currently running their scams on layer one are launching a "new" token, something that has been done before, only this time on Bitcoin. Only a complete ****** totally bereft of imagination thinks that this is innovation, or a good thing, or useful in any way. They can't conceive of a world where building on bitcoin is like building on the web. It's beyond their power to mentally process and sort. But this is where you live, in 2026. Ossification and client proliferation will keep bitcoin clean, force all low IQ, low imagination, imitative, Cargo Cult, mentally deficient, estrogenized, quote ********* manlets from despoiling the Golden Path of Bitcoin. It will allow a plethora of new specialist clients to emerge, enabling every "use case" anyone can conjure. Hope this helps!

• 
  Tech Ultimatum (@TechieUltimatum) reported 58 minutes ago

  OpenAI just made app building as easy as writing a sentence. With the new Sites feature in Codex, users can now create and deploy websites or apps from a single prompt. And ChatGPT now has 1 Billion users weekly. What it can do: • Turn ideas, docs, or notes into working apps • Deploy instantly with a shareable URL • Built-in authentication • Cloudflare D1 + R2 database support • Connect to tools like Slack, Google Calendar, and Google Drive • Create dashboards, internal tools, event planners, customer portals, and more The bigger story: AI is moving beyond helping people write code. It's now helping people ship complete products. OpenAI says Codex has already reached 5 million weekly users, with knowledge workers adopting it 3x faster than traditional software developers. We're getting closer to a world where: "Build me an app for this" becomes the new "Can someone code this for me?" 🤯

• 
  Thom (@Thom_K_NL) reported 1 hour ago

  @kinngh @thomasgauvin @Cloudflare I need this.... Bad