Cloudflare Outage Takes Down Parts of the Internet | Tech Field Day News Rundown: November 19, 2025

Agents for cloud migration. Commvault Unity has arrived hero for Code. Arista and Palo Alto are gonna team up.

We're gonna be talking more about Microsoft and Anthropic. There's some AI driven espionage out there. And, uh, wait a minute, we're gonna see if cloud flares back online in the closer look in this episode of the tech field.

A rundown. Hello everyone. Welcome to the tech field, a rundown for November the 19th.

Uh, today is a day that I absolutely cannot endorse It's national play Monopoly day. You're probably thinking to yourself, oh yeah, my family's gotten into some fights before. I once played Monopoly with lawyers and they made me sign contracts, so never again.

Thankfully, I've pivoted that business into talking all about the tech news that's been going on this week. And, uh, joining me is my somewhat sleepy co-host, Mr. Alistair Cook.

Al, it was great to see you last week in person, but I think that Jet lags starting to catch up with you With, you know, it's one of the amazing things is, uh, being in the states after working with the team, I got to go to Ohio with the team and, uh, be in person with people. But I also got to experience it's Sunday in the, in the US while it's Monday in New Zealand. And I'm usually on the other side of that.

It was a fun experience, but yeah, it's a long way home. It is a long way home. But thankfully a lot happened in the time that you were gone on that plane.

And we're gonna be talking all about it because Yeah, well, you know, how often do you break the internet? We're gonna start off with a story about the last people that broke the internet, and that's our friends over at Amazon Web Services. Because AWS introduced a new AI powered agent that dramatically accelerates cloud migration projects by automating tasks that used to take us weeks.

The professional services delivery agent can generate proposals from diagrams or notes, deploy sub-agents for coding and testing and leverage AWS transformed to modernize legacy systems. It's all backed by insights from thousands of prior migration. These tools position AWS alongside other heavyweights like Google and Microsoft, and using AI to streamline large scale cloud modernization.

Al, can Amazon finally turn the corner and make it easy to move things onto AWS? I hope so, because historically we've picked up what we had on premises shoved into, into AWS or another cloud platform and been terrified at the costs that we got after a little while. So going through a good consulting process to do that migration to, to get the most value out of getting your applications onto the public cloud, that's, that's a really vital part of cloud adoption.

And the aim with this particular component, the AWS Professional Services delivery agent, is to shorten the cycle for that instead of taking six months to a year to go through that planning and discovery and specification to have some AI agents deal with the huge volumes of data that you get as you're going through that, that process and to make sense of things through here. So, doing some of the, the design and iteration a around what does my infrastructure look like on AWS that's different from how it looks on premises. What things should I rewrite in order to use new platforms?

You know, the, how many s are we up to 7, 8, 9 rs of migration that's, uh, that we get with AWS. Having some of those guidance and decisions is, is absolutely vital. And I've seen this previously that it looks very simple to move to a cloud platform.

You just take what you got and you shift it across. But there's so many interdependencies and discovering and analyzing those interdependencies, working out your cycles to, to move. That's actually really hard work.

And it does seem like it's a combinations of choices, uh, process that really does suit an AI agent. Will this lead to more successful and and faster migrations from on-premises to AWS remains to be seen. And it's a, it's a complex kind of process to go through.

And one of my concerns is that as you go through that process, you need to build knowledge within your organization about how to, how to actually enact the transition from on-premises up to AWS and the differences. My concern is that these agents might prevent some of that knowledge gain, that, um, internal awareness that's a vital to speed up a, a normal migration. Of course, it may remove the need for that 'cause the AI agents might be that smart.

Commvault, our good friends at Commvault have just launched Cloud Unity and AI powered platform that unifies data protection, recovery, and identity security across cloud and hybrid environments, tackles AI driven data sprawl and rising identity based attacks, uh, with integrated governance tools and all of the nice advanced threat detection and recovery testing. Got access is available now, but full releases coming early next year. Tom, have you had access to the early, I have not had early access, but I was in the audience for last week's keynote when Sanjay Meti came out and kind of start talking about what Unity entails.

Uh, we're gonna be, uh, following along with the live stream that's actually happening today on the 19th. com for more details on that. Here were some of my big takeaways from what I got during the Commvault Shift keynote.

The first thing is res ops. I hope you're ready for a new ops, uh, designation. Uh, in this case, you know, with Commvault, it's resilience ops, right?

It is not enough to be available. You have to be resilient. You have to make sure your data never goes down.

Uh, that is part and parcel of what Commvault has always been about. We know that for a fact because we've seen that over the years. Commvault's kinda the gold standard for backup and recovery, but that's not the market that we're in anymore.

We are in the data protection market. And one of the things that they announced that, it took me a minute to kind of understand the totality of it, is something called synthetic restores. And you're probably thinking to yourself, well, how hard is it to restore data?

But I want you to think about this because this is the way that the market has been going. So it used to be that if something blew up, we had to restore it, right? Well, we had to go back and figure out where we had good data from.

And in the case of a security incident, we have to make sure that the data is still clean, and that means restoring back to, um, a certain point in time, right? But how do we know that the point in time is far enough back that we're getting the right data to restore? And what happened to all the data that's been created since that restore point?

Can't all be bad, right? So what Convault is proposing with this synthetic Restore capability is that they're gonna take a look at the backup. So we're not just going to last yesterday or last week or last month.

We're gonna take all of the backups and we're gonna look for the critical files that are infected, and we're gonna roll those back to the point where they were not infected. But for everything else that's not destroyed, we are going to roll it back to the last known good of that file. So if that last known good file was yesterday, then we could restore part of the system to a month ago and another part to last night.

That means minimal data loss, that means resilience all over the place. And that is huge for people who are uncomfortable with this idea of like, you know, we gotta start taking big full backups and full snapshots all the time because you never know when we're gonna miss something and we're gonna have to roll back. The other thing that I thought was big deal was the fact that Convault is integrating identity protection into their platform.

'cause one of the things I think that a lot of people are missing out on is that you can get infected and you can clean up all of the servers, but if the attackers penetrated into active directory, they've got a foothold, they can just keep coming back. Like, what if they changed the password on backup operators and added themselves as a unknown user, or worse yet at one that looks like a system account, they can just keep coding back in and doing everything they want to do, and you're just gonna have to keep rolling back. Only now they're watching you roll back in real time and they're combating all of those rollback points.

So eventually you're either gonna be forced to pay or you're going to be extorted for whatever else and have your data dumped. So I like where Convault's headed with this. Uh, the, the name change of course is important because we're unifying resilience into the platform.

And, and I'll admit some of the commercials in the keynote we're kind of cute. I kind of liked it. Uh, so make sure you stay tuned for more of that.

And also we've got some great content coming out from the people that attended, not my just myself, but Jack Poller and, uh, Jay Krell and More Great Field Day delegates. So make sure you're tuned in for that. We're going back to AWS because they also have something else they came out with.

It's their Kero AI coding tool that has new features aimed at producing more reliable, secure, and testable code, including everybody's favorite, the CLI version. It also has proprietary based testing to validate behavior against specifications and the ability to rewind to earlier checkpoints. The update also enables Kiro to work across multiple project routes and support custom AI agents while qualifying startups can get a free year of Qro Plus Pro Qro.

Pro plus AWS engineers say that these enhancements promote a more disciplined, specification driven workflow that reduces debugging stress and scales AI power development with quality rather than speed. Al I kind of like the vibes on this one. Is Kero gonna turn the corner from letting AI just kind of write whatever looks good?

Well, I think that's fundamentally what Kero ISS all about is not having those vibes, um, moving to a, a position where you write the specification of how your application should work and the AI implements what's written in that spec. Uh, I haven't gone hands on with this. The preview was, was launched in July and, uh, has has got a bit of buzz around.

I've, I've heard a little of, of people using, and I say, I really like the idea that there is a more declarative way of working here. This is define at the beginning how the software should work and then validate that it does work. Uh, in some ways it's, it follows a little bit of, of what I liked in the test driven design methodology where you first test for the functionality that you're going to build and make sure it fails 'cause you haven't yet built the functionality and then you build functionality until you, uh, pass the test.

Uh, that is one of the ways of going around. Let's design how things should work and, and make sure they do work that way. Uh, I definitely, this, this testing by properties is a really good functionality in here as well.

This is the ability to say this particular function should have this, this result. Now build a set of tests that, uh, uh, somebody walks into a bar and orders a beer, orders 99 beers, orders minus one beer, uh, orders a car. Uh, does the software do what it's supposed to do?

That, that kind of building those use cases, uh, those tests. So unlike test, different development where you build those first, the property based test approach builds those from your specification of how things should work. I like that whole declarative behavior.

Uh, and then the ability to work with larger projects, integrating Kero into your pipelines by using that CLI rather than having to do everything through a, uh, a, a click ops. Absolutely. This, this kind of stuff is where we're gonna see value in actually building real applications that are supportable and maintainable in the long term.

Yet don't require the vast number of developers hacking on little bits of code and particularly repeated bits of code over time, uh, remains to be seen how the pricing works versus value on this. And whether AWS can continue to deliver this, uh, a agent ai, almost a agent ai, um, for, for the long term, uh, there's, there's a lot still to be written about the cost of delivering these things versus what you can charge people for them. So we'll keep an eye on it.

Over time, Arista and Palo Alto Networks are expanding their partnership to address the rising hybrid data center complexity. And of course, AI driven cyber threats. Uh, integrating Arista's, uh, collection of product di Ava MSS with Palo Alto's own Next Gen Firewall and, uh, Prisma Air.

Uh, they offer a unified zero trust, uh, security and realtime threat. Quarantining centralized policy management, always my favorite. And DevOps friendly automation, delivering scalable, consistent protection across multiple data center and multi-cloud platform, uh, environments.

Uh, this sounds like there's a partnership between uneasy friends and that hopefully this is gonna be magically wonderful for customers. It should be wonderful for customers because when you look at what people are doing, especially in the cloud environment, they're leveraging a lot of Arista switches. That's where Arista has made a lot of their money over the last few years.

But one of the things that Arista is having trouble getting off the ground is a security practice. They've purchased a number of security firms over the last few years. I don't necessarily know that they've really taken off as much as we might like.

And so what do you do when you can't really build it? Well, you gotta buy something. And in this case, the buying is involved in a partnership, but they picked the best one on the block, right?

Palo Alto Networks is one of the, if not the premier security firms in the business, right? They have next gen firewall. They have a lot of cloud security offerings.

What they don't have is an inroad into those clouds through the networking team. Because as more and more people are starting to recognize the fact that networking security are two sides of the same coin, how do you displace existing installations? Well, in a lot of cases you have to partner up with a company who's displacing other things out of the network.

Arista made their money displacing Cisco. So how do you get into that network now that Arista has displaced them? Well, you partner up with Arista and Arista says, well, our stuff works really well with Palo Alto.

If you've got a refresh coming up, now's an opportunity to take a look at it. I will say though, that when this announcement was made, there were a lot of eyebrows that got raised because historically these kinds of partnerships lead to more engagement down the road. I'm not going to say the a word because that would be speculative at this point, but if there was a transaction to be had, this would not be a bad place for it to occur because both of these companies are very well known in their individual spaces, and I think wrapping them up would be a very unique opportunity for some people.

I don't think that that's gonna happen anytime soon though, because while Palo Alto is sitting on a lot of money, wrist is worth a lot. And I don't think that this is an acquisition that they could really swallow in whole yet, but it would be very complimentary. And, uh, quite honestly, I, I'd I'd love to see, uh, someone like Akin do to doing some Palo Alto presentations, because that could be real fun.

Microsoft announced yet another major new AI partnership with Anthropic and Nvidia, but this is signaling a move beyond the once exclusive ties that it had to open. AI and philanthropic will buy $30 billion in Azure compute while invest. Nvidia invests up to $10 billion and Microsoft up to $5 billion to help scale Anthropics Claw AI models.

This deal gives anthropic massive Nvidia powered capacity and strengthens Microsoft's AI infrastructure strategy as its relationship with open AI involves at Microsoft Ignite, which is happening this week, they also introduced a new IQ lineup of AI tools, work iq, fabric IQ, and Foundry iq, as well as Microsoft Agent 365, designed to automatically support business workflows as companies prepare for future of billions of AI agents. Al, I think it's interesting that we've always heard so much about the partnership that Microsoft has with OpenAI, and now they're going out and seeking out Anthropic. Do you think Microsoft's playing the long game by betting on different horses, or is there something else going on here?

I think this is definitely a, a hedging strategy. So we covered previously on the rundown that OpenAI pre had had an exclusive deal with Microsoft that OpenAI would run on Azure and that Microsoft would use OpenAI as its standard AI powers, things like copilot. We covered that.

This tight deal had come to an end and there was a much lucid deal where OpenAI could use other vendors and were contracting to use other vendors to provide that infrastructure and that Microsoft is free to work with other vendors. So Microsoft has always been the king of partnerships partner with everybody. Uh, I see this relationship with, uh, anthropic as being part of that.

They recognizing that there is no longer an exclusive relationship. They need to have a real, real relationship with a competitor to open ai in this case anthropic, and it's all part of the same partner with everybody and make sure that nobody can hold you to ransom. So OpenAI can't say to Microsoft, well the, this is the only feature set you get because this is what we're, we're doing.

Uh, Microsoft has then no leverage to say, well, we want this other feature we need for copilot as they make a commitment into Anthropic. Yeah, it absolutely gives them some more choices in there, uh, naturally enough, this is the AI money go round, and so Nvidia is in here as well. And Nvidia investment in anthropic means that Anthropic will hand back some of that money to buy Nvidia hardware.

So when we talk about this as a money go round, it really is money changing hands back and forth in some of these places. Um, Microsoft continues to, uh, put money into this, uh, $5 billion, uh, invested here. So it's not a trivial amount, but it's not the $30 billion, $60 billion we've seen in in other deals along the way.

Um, and Microsoft has $13 billion in open AI investment. Uh, it really is, I think, a, a protection against, uh, open AI choosing their own path. We've seen some interesting challenges with open AI as they wanna change their, their governance and structure around there.

Uh, if that represents a risk to Microsoft, connecting up with Anthropic seems like a great way to mitigate that risk. Uh, we'll see just over time as, uh, whether this continues to grow, we see more and more money being put into, uh, anthropic and maybe a, a dial off on OpenAI. I don't think so.

I think this is really a protection of the relationship with OpenAI. Newly uncovered cyber attack shows the first large scale espionage campaign carried out mostly by ai. A apparently Chinese stake backed group used AI agents to handle up to 90% of the operation, so including scouting potential, uh, victims and exploiting vulnerabilities in stealing data.

Uh, the incident highlights the rapid escalation of AI driven threats and the need for strongest safeguards around these AI attacks, better detection and also industry-wide cooperation in as near real time as possible to protect against these attacks. Tom, uh, this is another arms race, isn't it? It's AI for attack and AI for for defense, Yeah.

And that's exactly what we talked about on this week's episode of Security Boulevard, because it really did feel like this is kind of turning the weapons back on the creators a little bit. I also thought it was a really fascinating way that they were able to kind of slice this attack up so that multiple Claude agents were not actually knowing what was going on. It was almost like a operating in a compartmentalized cell structure so that each of the agents were returning work.

It was then being tied together by other agents to do the actual exploiting so that it evaded all of the models jailbreak capabilities of saying, oh yeah, we're not gonna do bad things. You know, theoretically it'd be like, you know, the difference of me asking how do I rob a bank versus describe what good bank security looks like? Uh, one of those things would probably set off a trigger while the other one when used improperly would probably get me the answers that I need to know how to avoid security cameras and guards and things like that.

Uh, also, uh, the, the press release from Claude, uh, I'm sorry, from philanthropic about Claude was rather interesting in the fact that they said, well, we detected it and we stopped almost all of it, but almost all of it ain't all of it. And it, some of the attacks did go out, and of course now they're probably going behind the scenes to try to figure out exactly what was used to bypass the filters and the protections and how they're going to evade it in the future. And, uh, I I think that we're just, we're starting to see the tip of the iceberg here because this is probably not the first one that we've seen, but it is definitely the first one that somebody's reporting about.

So, bravo to philanthropic for at least admitting that this was going on. But I think one of the things that we're gonna see as this goes on more and more is that people are gonna refine their prompt engineering. They're gonna be able to break these tasks down into finer and finer detail so that it really is going to become impossible or worse, yet they're gonna take the outputs from one group of AI and feed it to a separate set to actually leverage the attack.

So that attribution is gonna be very difficult to trace down. And I, I don't know if there's a clear cut answer here, because the real thing that people are saying we need to do is lock it all down so that it can't be used for attack. That's like saying, oh, well, we should get rid of all VCRs because they're only ever used to record, you know, copyrighted programs when that's actually not the case.

Uh, we just, we have to find a better way to figure out to use them. All right, we wanted to take a closer look at a story today. Uh, it took us a little bit while to write it, of course, because there were some internet issues.

Everyone's favorite web infrastructure provider, CloudFlare had a massive global disruption that caused error 500 messages and took down major platforms including X Twitter and chat GPT along with most other AI platforms. The outage, uh, triggered by a sudden spike in unusual networking traffic impacted thousands of websites early on Tuesday morning and highlighted the fragility of internet architecture. 20% of global web traffic flows directly through CloudFlare.

The firm is actively investigating the root cause and working to restore full service while underscoring the need for greater resilience and digital infrastructure. And I think it's kind of funny that we've run into this problem, uh, recently, al where DNS update and AWS took down one half of the internet, and CloudFlare getting knocked offline by a massive traffic spike, uh, took out part of the rest of it this week. Uh, you know, they, they're still attributing what exactly went on.

We know that CloudFlare is gonna give us a great postmortem when they figure out exactly what happened, but I want to turn it over to you and maybe ask, are we putting too many eggs in one basket by relying on CloudFlare to protect us? Because when CloudFlare goes down, we can't get anywhere. So here's the thing.

CloudFlare is very popular because it does a great job of a very specific task of delivering applications globally, uh, to your users in a way that gives them really good performance anywhere in the world. CloudFlare has built an amazing network to do this. And then let's, let's put this in context.

Although we've centralized everything on CloudFlare, we don't often see CloudFlare outages. We certainly don't often see CloudFlare outages of this scale. So yeah, the outage is extremely visible the same way any of these centralization outages are extremely visible, but they're very rare.

And so when you look at what is the impact on the total uptime of whatever system you're using, you'll still find that even with the, this particular failure or this type of failure, uh, you're still getting higher availability and better performance for your users across the, the last year than you would've if you ran this yourself or you tried to do some equivalent of this. So, yeah, centralization means it's very visible when things go wrong, but centralization means you can spend a lot of engineering on making sure they don't go wrong often. One of the things I've read on this, and, and it came across Reuters, is that the, the sort of triggering event here is an automatically generated config file that got too large and then caused a service to crash.

Um, this feels a little bit like what we saw with CrowdStrike where a config file was deployed out and that caused the massive outage for, for, um, CrowdStrike. Uh, it seems like these config file changes need to run through a CI CD pipeline to validate they're not gonna cause problems, yet we need to deploy them out fast. You're the eternal optimist, al, and that's what I love about you because I, I took a, a slightly different, uh, tack from this and it come courtesy of our friends at down detector, you know, the website that everybody goes to, to figure out when you're having problems with stuff.

Oh, wait, I couldn't check that one either. 'cause it turns out it runs on CloudFlare. Yes.

I, the world is better off because CloudFlare is front-ending sites and preventing massive DDoS attacks. And their engineering has done a really great job of helping us extend and expand the way that that web works. And, and a lot of other things too.

They're effectively the Internet's proxy layer at this point. 1 from the trash heap of history, uh, because of bad configuration stuff. All that being said, you cannot put all of your eggs in one basket.

And, and we've learned that from a lot of things. I mean, all you gotta do is just look up the infographic from the XKCD of, you know, the entire modern internet. And then there's a little thing down there, a little Jenga piece, and it's insert name there, AWS CloudFlare, uh, NGINX, whatever.

What we're starting to see is that when these outages do happen, they happen at a a, a level that is difficult to contain. And you mentioned CrowdStrike. We've talked about AWS before.

Um, there are a lot of challenges that happen when we've built something that is effectively too big to fail and more than any other company, I think CloudFlare is pretty strict in the way that they deploy things, in the way that they, uh, leverage stuff. I can't fault them for this. How can you figure out, oh, the config file got too big.

When do we test for that? How do we understand that the problem is not the outage itself, it's the rolling effects from that outage. If, you know, it went down for 10 minutes, everything that checks on CloudFlare, it kept pinging, it kept going offline.

You know, little things that you didn't think were reliant on that. I was having a meeting first thing this morning and went to go check the calendar on somebody's website. It's offline, it's hosted on CloudFlare.

You're like, oh, well that's fascinating. And so you've got to figure out how to prevent this from happening. Yeah, in some cases, you, you probably do need to have your, your load balanced infrastructure running on it, but I don't know, maybe put your status page somewhere else, hosted somewhere that nobody goes like Oracle Cloud.

So you're, you're suggesting to avoid the, the problem that AWS had along with outages when they had the big S3 outage and the, uh, the website at AWS that reports on the status of AWS services is dependent on the S3 service. Uh, yes, circular dependencies like that are really problematic, but, uh, yeah, I, I still view using a a well-engineered system that is designed to continue to operate at large scale absolutely is, is way better than any of the other solutions for this. So yeah, I, I'm still comfortable with putting our websites through CloudFlare, having even trivial things go, go through CloudFlare.

But yeah, the tool that tells me whether things are working or not can't tell me if they're not working, if it's dependent on the thing that's not working. Yeah, mindful of circular dependencies. The other thing that strikes me on this is that there's, there do seem to be a lot of circular dependencies that go through CloudFlare, because what we didn't see is when the, the issue was resolved, everything miraculously started working again straight away.

It took a little while for it to filter through the various layers that are dependent on CloudFlare before all of the layers caught up with one another. You know, this eventually consistent, uh, systems that we usually see at internet scales, uh, would be interesting to see if you had centralized everything into a single place, rather than having these eventually consistent distributed systems, whether they time to get back into operation would've been longer or shorter, because of course, in a centralized place, you have to replay everything in a single stream. Whereas when you're doing eventually consistent, you're replaying in different locations and your total timeout might be lower.

Something that's not going to be affected, I hope by any cloud outages, any security outages. Um, now I'm gonna have to make some sacrifices To all of the good luck Gods, something that I expect not to be, uh, affected is AI infrastructure field day four. That will be my next return to the United States January 28th and 29th.

The event is already filling up. We already have, uh, four companies confirmed to be there, and we're expecting a few more to roll in as well. Uh, really looking forward to kicking off 2026 for Tech Field Day with AI infrastructure field day.

Uh, and then of course, Tom, you're going a long way, a field as well. That's right. We're looking at the possibility of heading over to Cisco Live and Maya, which is gonna be in Amsterdam once again this year.

You know, I can't get enough of those little teeny tiny pancakes. com is gonna be, uh, your home for that. So when Al flies halfway across the world, I fly the other halfway across the world.

But then Al you're, you're coming back in March? Absolutely. I can't stay away.

Uh, I'll be back for Cloud Field day 25 in, uh, the middle of March. Uh, and those tiny, uh, pancakes, the puffs, uh, my Dutch, uh, sister-in-law has, uh, gave gifted us the pan to make them. So maybe come down, visit me and we'll make you some tiny pancakes.

Uh, what isn't tiny, of course, is the tech Field Day rundown. Do join us, uh, continue to join us for the Tech Field Day Rundown. You can catch new episodes every Wednesday, either as a YouTube video or in your favorite podcast application.

Rundown streamed on Techstrong tv, of course, that's part of the Futurum Group. And you can find us on both Techstrong and RUM Group, uh, locations. We'll be back next Wednesday to talk about all of the IT news for the week.

That was, and until then, for myself and for Tom Hollingsworth, and for Corey and all of us here at the Field Day team, we're wishing you and yours a great day. Great week, and we'll see you on Wednesday.