That would be idiotic not just because unrusted parties can use those addresses, but more importantly because those are more or less terrible hacks that should be avoided completely if possible. You rather should have globally unique addresses on your internal network if you can, which would just break this.

> Perhaps a rather drastic change like this will lead to more user friendly ways to install self-signed certificates on home networks.

That's also not sensible. The whole idea of linking stuff to specific networks is bad. There is no reason why access to a device on your home network should in any way be linked to your client device being connected to that same network. It's the internet, not "the home network and the cloud".

What is needed is a way to establish a trust relationship between two devices that you have control over. Where those devices happen to be connected to the internet should be absolutely irrelevant. There might be an argument to be had to maybe support for a simplified peering procedure on a local network might be a good idea--but the point is that once the trust trelationship is established, you should be able to move your client device to a different network on the other side of the planet and still be able to talk to your device on your home network.

Seems quite unavoidable with IPv4, or what is the hack you're referring to specifically?

> That's also not sensible. The whole idea of linking stuff to specific networks is bad. There is no reason why access to a device on your home network should in any way be linked to your client device being connected to that same network.

What is the reason against? It's not like the idea of a privately managed network in which you trust all peers is novel or rare. Most people have a network in their home that they manage for themselves or for their family. It's the perfect scope for IoT devices.

> It's the internet, not "the home network and the cloud".

That's an interesting notion, but unfortunately does not reflect reality of use or the design of existing internet protocols, or even the very core concept of the internet: interconnected networks. I have a private network at home. My means to connect devices on this network to the internet is via a gateway which is assigned a single globally unique internet address by my service provider, and a locally unique address on the private network.

If at some point every device has its own global address and is accessible globally, it will be more accurate to assume that something is insecure if it communicates in the plain, but we're not there yet. What the browser is doing now is pretty much assuming an arbitrary level of "better safe than sorry".

> What is needed is a way to establish a trust relationship between two devices that you have control over.

Say, by sharing sharing keys over a network under your control, certified and authorized by a device you trust for pretty much everything else on that network?

> Where those devices happen to be connected to the internet should be absolutely irrelevant.

Agreed, but the current conundrum is that they need to be connected to the internet if you want to use a central certificate authority.

Well, yes, for most people it unfortunately is. But imagine if your are one of the lucky ones who do have global IPv4 addresses everywhere. And now someone sells you a product that tells you "sorry, nice IPv4 network that you have there, but you have to install NAT and an RFC1918 network to use this IPv4 product". Not very sensible, is it? Same applies for IPv6 and ULA, obviously.

> What is the reason against?

What would be a reason for limiting the usefulness of your devices?

> It's not like the idea of a privately managed network in which you trust all peers is novel or rare.

Which is fine, but not a sensible assumption to make in an IP product. If you want to use it in a privately managed, trusted network, of course you should be able to, but the idea that an IP device should just refuse to work over IP if your IP happens to extend beyond your LAN is idiotic. That should be a matter of the network's policy, not of the device's hard-coded policy.

> That's an interesting notion, but unfortunately does not reflect reality of use or the design of existing internet protocols, or even the very core concept of the internet: interconnected networks.

Erm ... that's completely backwards? It unfortunately does not reflect the current use of IPv4 in particular due to NAT everywhere, but that certainly was not part of "the design of existing internet protocols", that was a hack due to lack of addresses.

What was before the internet were separate local (and sometimes not so local) networks: You had all kinds of link-layer protocols, and then various higher-level protocols, usually specific to a given link-layer. The whole point of the internet was to add a common abstraction to all of those link layers protocols, precisely to eliminate any distinction between local or remote, ethernet or token ring, modem or ISDN, GSM or CDMA, an addressing layer that erased the distinction: If you had an IP address and the thing you wanted to communicate with had an IP address, you could, even if you were on token ring, your WAN link was ISDN, the backbone was ATM, the peer's WAN link was a dial-in modem and their LAN was ethernet. The point of IP is that you don't have to care, any IP address is as good as any other.

> I have a private network at home. My means to connect devices on this network to the internet is via a gateway which is assigned a single globally unique internet address by my service provider, and a locally unique address on the private network.

Well, yes, unfortunately, that is the case nowadays. That is not how IP was meant to be used, and it's causing massive problems. If it weren't for lack of addresses, your home network should have a globally unique /24 or something (and it did, back in the day).

> If at some point every device has its own global address and is accessible globally, it will be more accurate to assume that something is insecure if it communicates in the plain, but we're not there yet. What the browser is doing now is pretty much assuming an arbitrary level of "better safe than sorry".

Not sure I am getting your point!?

> Say, by sharing sharing keys over a network under your control, certified and authorized by a device you trust for pretty much everything else on that network?

Well, arguably you totally should not trust your router, they tend to be crap security-wise.

But in any case, my point was that at most that should be a pairing mechanism. So, once the trust relationship is established, there should be no need to stay on the local network for further secure communication.

> Agreed, but the current conundrum is that they need to be connected to the internet if you want to use a central certificate authority.

Well, yes?! But the solution is not to hard-code policies that prevent full use of IP.

Are you arguing from the assumption that my suggestions and any other form of establishing trust are mutually exclusive? If you're that lucky guy with a global address for your lightbulb, by all means use what's at your disposal to establish a trusted encrypted link between the device and the user in a convenient way. Not sure how that would prevent the vast majority using these on private networks with a different method of authentication and different criteria for trust.

> Erm ... that's completely backwards? It unfortunately does not reflect the current use of IPv4 in particular due to NAT everywhere, but that certainly was not part of "the design of existing internet protocols", that was a hack due to lack of addresses.

So, given the limited address range, it was clearly not designed for every person in the world to have an address, not to mention every appliance in your kitchen. The internet has grown rather organically and has adopted a broader use case. The infrastructure, protocols and best practices used on the internet now reflect this unanticipated use case.

> Well, yes, unfortunately, that is the case nowadays. That is not how IP was meant to be used, and it's causing massive problems. If it weren't for lack of addresses, your home network should have a globally unique /24 or something (and it did, back in the day).

How it was meant to be used is an artefact that stopped mattering some time in the 80s.

> Not sure I am getting your point!?

The point is that flagging plain http websites as "unsafe" makes a lot of assumptions about my network. They're not necessarily unsafe. In one case, it's on my apartment-wide LAN. In another case, it's connected by ethernet directly to the client. Neither of these are particularly exotic topologies.

> But in any case, my point was that at most that should be a pairing mechanism. So, once the trust relationship is established, there should be no need to stay on the local network for further secure communication.

Why not both?

> Well, yes?! But the solution is not to hard-code policies that prevent full use of IP.

Agreed? I'm not sure where you got the idea that I think that any of these things should prevent the full use of IP. Certainly not from anything I've said.

Well, if that were a standardized way to establish trust, that necessarily would lead to vendors adopting it at the cost of suporting other kinds of setups?

Also, it is very problematic to overload not globally routable addresses (also often misleadingly called "private addresses") with security semantics. While many home setups do have a sort-of security boundaries around RFC1918 subnets, there is absolutely no guarantee that that is the case. So not only would such a mechanism break "sane" (i.e., NAT-free) setups, it also would make otherwise perfectly fine and useful setups risky. Have a VPN link to another company that also uses RFC1918 space or ULA, and suddenly your IoT stuff starts trusting that other company. Or even just if you happen to have departments that aren't supposed to trust each other, and that happen to have a common ULA prefix, and now some devices simply assume trust where none is implied/make it impossible to use otherwise perfectly fine setups because of unjustified trust assumptions. Or simply a guest on your network. Or ... whatever else that can share non-globally routed address space with you without any trust implied.

> So, given the limited address range, it was clearly not designed for every person in the world to have an address, not to mention every appliance in your kitchen. The internet has grown rather organically and has adopted a broader use case.

Well, yes, but that wasn't because it was intended to be used with NAT, or anything else that was not a globally routable address for every device, but because it wasn't expected to gain that many users.

> The infrastructure, protocols and best practices used on the internet now reflect this unanticipated use case.

Which is an argument for what exactly? Especially with regards to IPv6 and ULA?

> How it was meant to be used is an artefact that stopped mattering some time in the 80s.

Why would that have stopped then? Again, in particular with regards to IPv6, which does not have the address scarcity that might have justified use of NAT and non-globally routed address space as a temporary workaround?

> The point is that flagging plain http websites as "unsafe" makes a lot of assumptions about my network.

... just as not doing so does? If anything, it would be arbitrary to just exempt certain prefixes from security policies when there is not normative basis for such an exemption. I happen to have only globally routable IPv6 addresses on my LAN, but my LAN is indeed trusted, both wired and WiFi, using distinct /64s. But I also have a guest WiFi that is in the same IPv6 /48, which is not trusted. And I have a VPN link to a customer of mine that uses RFC1918 address space, which is absolutely not trusted.

So, yes, it is "better safe than not safe". But it's exactly the opposite of arbitrary, in that it does not make any assumptions about your network, it provides security no matter what the details of your network, and using the exact same policy for everything. And it's hardly "better safe than sorry", given that this is all a result of being very sorry about all the crap that resulted from lack of security so far.

> They're not necessarily unsafe. In one case, it's on my apartment-wide LAN. In another case, it's connected by ethernet directly to the client. Neither of these are particularly exotic topologies.

Yeah, and how is your browser supposed to know that?

> Why not both?

What both?!

> Agreed? I'm not sure where you got the idea that I think that any of these things should prevent the full use of IP. Certainly not from anything I've said.

The question is not whether it should, but whether it would. Suppose browsers were to implement a policy of "RFC1918 and ULA are considered safe unencrypted and -authenticated". What would vendors of devices do? I guess we can agree that they would use that policy for config access, as it simplifies the design of their devices, right? Now, that would cover 99%+ of their current user base. Which probably means they won't bother providing an alternative mechanism. Which means (a) you can't use their devices in other setups and (b) their users are locked into such setups, which makes it impossible for, say, router vendors, to build more useful networking products that use the full potential of IP.

So on one hand you believe that supporting one standard will necessarily come at the cost of supporting another (I don't), and you agree that what I suggested might cover 99%+ of the current user base, yet you favor a solution that depends on global addresses, something which definitely doesn't come close to 99% of potential users?

> Also, it is very problematic to overload not globally routable addresses (also often misleadingly called "private addresses") with security semantics.

Well, if you want to be really anal about it you could call them "addresses which fall into one of the address ranges allocated for private use", but you're just splitting hairs.

> While many home setups do have a sort-of security boundaries around RFC1918 subnets, there is absolutely no guarantee that that is the case.

There is no guarantee, but that's different from saying that it's inherently unsafe.

> Which is an argument for what exactly? Especially with regards to IPv6 and ULA?

It's a reflection on how the internet is built. It doesn't matter that in the ideal network, everything might have an address, when pretty much every device is behind some kind of NAT. IPv6? Come back when it's widely adopted.

> Why would that have stopped then? Again, in particular with regards to IPv6, which does not have the address scarcity that might have justified use of NAT and non-globally routed address space as a temporary workaround?

It stopped mattering because of address exhaustion and slow adoption of IPv6. Now, NAT is an integral part of the internet. I'm not trying to justify it or state this as a matter of preference—I'd definitely prefer having a ton of IPv6 addresses over the single IPv4 address I actually have—I'm just laying things out as they are, and how they are for the vast majority of consumers.

> ... just as not doing so does?

No. Flagging a website as "safe" when it can not be established that it is safe is at least as wrong as flagging it as "unsafe" when it can not be established as being unsafe. What I'm suggesting, not flagging it in any particular way at all, would be taking a neutral stance. IMO, the practice of calling HTTPS sites "secure" is itself potentially misleading to consumers. It is only secure in a very specific sense, likely not in the broader sense a layman would consider.

> Yeah, and how is your browser supposed to know that?

The question I stop at is "why is my browser supposed to know that?"

> What both?!

Both a way of verifying and distributing certificates network-wide in a LAN and for those certificates to be usable globally.

> The question is not whether it should, but whether it would. Suppose browsers were to implement a policy of "RFC1918 and ULA are considered safe unencrypted and -authenticated". What would vendors of devices do?

The premise of my suggestion is that the browsers won't back down from indiscriminately marking plain HTTP sites as insecure, hence "Perhaps a rather drastic change like this will lead to more user friendly ways to install self-signed certificates on home networks."—so suppose they would support such a method. Would that be better or worse than current practice?

> Which probably means they won't bother providing an alternative mechanism. Which means (a) you can't use their devices in other setups and (b) their users are locked into such setups, which makes it impossible for, say, router vendors, to build more useful networking products that use the full potential of IP.

That's a load of conjecture. I'm not sure how to respond except with a bunch of other conjecture, so I'll refrain.

So, you think vendors who have covered 99%+ of their userbase with a solution will generally also implement an alternative that is way more complicated for the remaining 1%?

> and you agree that what I suggested might cover 99%+ of the current user base, yet you favor a solution that depends on global addresses, something which definitely doesn't come close to 99% of potential users?

No, I favor a solution that does not depend on the global (non-)routability of an address, i.e., a solution that works for 100% of users.

> Well, if you want to be really anal about it you could call them "addresses which fall into one of the address ranges allocated for private use", but you're just splitting hairs.

But that is still equally misleading. There is nothing "private" about those addresses, and in particular nothing "more private" than globally routable addresses. Anyone can use those addresses, all the RFC essentially says is that you won't collide with addresses allocated by RIRs, but they might collide with other administrative domains that choose to use the same prefix. That doesn't mean that you cannot use them on a WAN, or between companies, or really anywhere where you can agree with all participating networks on the allocations. All it means is you have to expect collisions if you connect previously separate administrative domains, and that you cannot expect yout ISP to announce them for you on the public internet, that's it.

Also, just as you can use non-globally routable addresses between networks, you can use globally routable addresses for private networks, and you should if you can (which in practice means when you build an IPv6 network): Even if you build a network that is not intended to be connected to the internet at all, if you do have a globally routable IPv6 prefix allocated for your organization, you should number that network with addresses from that prefix.

> There is no guarantee, but that's different from saying that it's inherently unsafe.

No, it's actually not. "unsafe" does not mean "you will hurt yourself", it means "it has not been established that you won't hurt yourself".

> It's a reflection on how the internet is built. It doesn't matter that in the ideal network, everything might have an address, when pretty much every device is behind some kind of NAT. IPv6? Come back when it's widely adopted.

So, for the question of how to achieve (as close as possible to) an ideal network, it doesn't matter what the ideal network would look like?! Or do you think we should just wait until device vendors have screwed up IPv6 before we try to enforce some sensible policy?

> It stopped mattering because of address exhaustion and slow adoption of IPv6. Now, NAT is an integral part of the internet. I'm not trying to justify it or state this as a matter of preference—I'd definitely prefer having a ton of IPv6 addresses over the single IPv4 address I actually have—I'm just laying things out as they are, and how they are for the vast majority of consumers.

So ... because noone uses IPv6, you suggested to use ULA as an indicator for security?! I am not sure I follow ...

> No. Flagging a website as "safe" when it can not be established that it is safe is at least as wrong as flagging it as "unsafe" when it can not be established as being unsafe.

No, you don't establish "unsafety", that is the default assumption. The only way to establish that something is unsafe is to show after the fact that someone got hurt, which is just completely useless as a security mechanism.

> What I'm suggesting, not flagging it in any particular way at all, would be taking a neutral stance.

Wouldn't a neutral stance be to instead display a security status of "security unknown" (which is obviously equivalent to insecure)? "Not flagging it in any particular way" simply means that the user makes an assumption one way or another, not that the user thinks "it is unknown whether this is secure".

> IMO, the practice of calling HTTPS sites "secure" is itself potentially misleading to consumers. It is only secure in a very specific sense, likely not in the broader sense a layman would consider.

Well, yeah, but that is not really relevant to the question about warning about an insecure situation. Just because there are insecure situations that you cannot warn about, does not mean that therefore warning about other insecure situations isn't useful. Really, it makes much more sense to warn about insecure situations (which means, situations not known to be secure against certain types of attacks deemed relevant in the respective context) than to in display anything that says "this is secure", as security is alway relative to specific attacks, not a global property.

> The question I stop at is "why is my browser supposed to know that?"

Because your browser should help you protect your personal data that you process using your browser? I mean, I don't think it should know that, it should just enforce the same encryption requirements everywhere, but you seem to disagree with that because there are networks where your personal data is secure without encryption--in which case, your browser would either have to give up the goal of protecting your personal data, or it would have to know about which parts of your network are secure without encryption.

> Both a way of verifying and distributing certificates network-wide in a LAN and for those certificates to be usable globally.

Well, that would be a pairing mechanism then?! (Which still should not overload global routability with security semantics.)

> The premise of my suggestion is that the browsers won't back down from indiscriminately marking plain HTTP sites as insecure, hence "Perhaps a rather drastic change like this will lead to more user friendly ways to install self-signed certificates on home networks."—so suppose they would support such a method. Would that be better or worse than current practice?

It depends on the mechanism? Yes, centralized certificate management for your own devices would be useful, but it should not in any way overload the routability of addresses. If you want to use the LAN as a semi-trusted key exchange mechanism, that probably should happen at the ethernet layer. Or maybe with a one-hop TTL on the IP layer. You have to detect whether you are on the same LAN, not whether you are using an RFC1918/ULA prefix, because not all LANs use RFC1918/ULA, and nor all RFC1918/ULA are limited to a LAN, let alone a trusted LAN.

> That's a load of conjecture. I'm not sure how to respond except with a bunch of other conjecture, so I'll refrain.

It is mostly an observation of what always happens in such situations.

Not necessarily, but the designer of a standard could take both use cases into account.

> No, I favor a solution that does not depend on the global (non-)routability of an address, i.e., a solution that works for 100% of users.

What, more precisely? Of course taking into consideration that 100% of users might not even have an internet connection. One thing I'd like is for devices to a their key signature printed on a sticker. Then I can verify the signature, log in and generate a new key and password, generate a certificate that I can install myself or sign up with a service like Let's Encrypt.

> But that is still equally misleading. There is nothing "private" about those addresses, and in particular nothing "more private" than globally routable addresses.

No, it's not misleading to say that they are allocated for private use. Your ISP drops connections to these addresses because they respect RFC1819 and don't route to the private address ranges. Even if they didn't, these address ranges are still allocated for private use, and your ISP is Wrong. They're only routable in the sense that IP would technically allow it, but the internet is not simply IP but a collection of standards and best practices.

And sure, "private" has a very broad meaning. A browser could very well flag a certificate that was distributed from a private address as such and let the user decide whether they trust that source.

> Also, just as you can use non-globally routable addresses between networks, you can use globally routable addresses for private networks, and you should if you can (which in practice means when you build an IPv6 network): Even if you build a network that is not intended to be connected to the internet at all, if you do have a globally routable IPv6 prefix allocated for your organization, you should number that network with addresses from that prefix.

Sure. But again, "in practice means when you build an IPv6 network", i.e. not a typical consumer, for how long more? In an enterprise there are already many different ways to solve the problem of authentication, certificate signing and encryption. Consider that the average internet user doesn't even have a registered domain name or a static IP allocation.

> No, it's actually not. "unsafe" does not mean "you will hurt yourself", it means "it has not been established that you won't hurt yourself".

So everything on the web should be flagged by the browser as unsafe? I don't know how the browser can ever safely establish that I won't hurt myself. "Unsafe" and "safe" are two sides of a subjective, blurry line, at best a reasonable assumption and at worst an arbitrary handwave.

IMO, the browser is taking what should be exact descriptions of the nature of the connection and water them down to vague, misleadingly simplified concepts. The browser could tell me that my connection to a site is unencrypted, that it is encrypted with an uncertified key, or that it's not encrypted, and when you click them they could show a help text describing what that means exactly, the possible consequences of using the service and details on the key and certificate if applicable. When I click the "Secure" badge in Chrome, I don't even get to see which CA signed it, or a public key.

"Secure" and "Insecure" mean just that, rather impossible things for a browser to verify, and something that a user unfamiliar with the underlying technology may interpret as an authoritative rating of the provider of the service as a whole, when in reality there are many more aspects to take into a count in deciding whether a site is secure or insecure.

> So, for the question of how to achieve (as close as possible to) an ideal network, it doesn't matter what the ideal network would look like?!

Well, it involves IPv6, we can start there. We're talking about a new security policy that a major browser seems to want to implement shortly, definitely much more shortly than full IPv6 rollout.

> Or do you think we should just wait until device vendors have screwed up IPv6 before we try to enforce some sensible policy?

This is a very loaded question, given that we still disagree on whether a solution that works well both for globally routable and NATed devices is possible.

> So ... because noone uses IPv6, you suggested to use ULA as an indicator for security?! I am not sure I follow ...

I never said that no one uses IPv6, so I agree that you don't follow.

> No, you don't establish "unsafety", that is the default assumption. The only way to establish that something is unsafe is to show after the fact that someone got hurt, which is just completely useless as a security mechanism.

Let's say that I see your ladder. It's broken, so I tell you that it's unsafe. Unreasonable assumption? You take it down and bring another ladder. I don't see it, but I tell you it's unsafe. You see it and can clearly say that it isn't. Is it unsafe? Is it reasonable for me to tell you that it is unsafe? "Unsafety" isn't the default assumption that a browser makes (and with regards to plain HTTP in particular still isn't in the version of Chrome I'm using).

> Wouldn't a neutral stance be to instead display a security status of "security unknown" (which is obviously equivalent to insecure)? "Not flagging it in any particular way" simply means that the user makes an assumption one way or another, not that the user thinks "it is unknown whether this is secure".

Maybe that's actually the better option. But no, "security unknown" in that sense is not equivalent to insecure. As an extreme, I could create a network with an Ethernet cable between two off-grid devices that I control in a faraday cage. On the other end of the extreme, someone could be tapping a cable far away from my computer and figure out what connections I make regardless of encrypted data. Somewhere in between the two extremes, close to the likely, the party that I establish a secure connection too could be sharing our communication with other parties.

> Because your browser should help you protect your personal data that you process using your browser?

We've already established that the browser can't know it, so is the browser a fundamentally flawed concept?

> Well, that would be a pairing mechanism then?! (Which still should not overload global routability with security semantics.)

Yes? The only part you seem to disagree with is the possibility of having a router in a local network (and yes, local networks exist) facilitate and streamline the exchange.

> If you want to use the LAN as a semi-trusted key exchange mechanism, that probably should happen at the ethernet layer.

So again, a perfect application for a router? The router is in a perfect position to verify that I am on its network.

> It is mostly an observation of what always happens in such situations.

Yes, as evident from the absolute lack of overlapping authentication and encryption standards...

> On the other end of the extreme, someone could be tapping a cable far away from my computer and figure out what connections I make regardless of encrypted data. Somewhere in between the two extremes, close to the likely, the party that I establish a secure connection too could be sharing our communication with other parties.

Which is in no way in conflict with saying "this is insecure". That is in conflict with saying "this is secure", because that implies "... against this specific set of threats", which is not understood by the average user. So, yes, I agree, browsers should generally avoid telling users that something "is secure", but it is perfectly fine to say "this is insecure".

> We've already established that the browser can't know it, so is the browser a fundamentally flawed concept?

No, it's just a subjective entity as all entities in the world are, and so it has to determine risks based on incomplete information, as all entities in the world have to. Also, it's not strictly true that it cannot know that, but it cannot know that without you telling it. It might well be possible to have an option where you could tell your browser "this set of addresses is safe to talk to unencrypted and unauthenticated".

> Yes? The only part you seem to disagree with is the possibility of having a router in a local network (and yes, local networks exist) facilitate and streamline the exchange.

No, I disagree primarily with overloading the semantics of "private addresses", and with mechanisms that only allow communication in a local network. "private addresses" is neither reliably indicative of nor a required property of "within the same local network".

But also, a mechanism that does not depend on being on the same local network for pairing would be preferable.

> So again, a perfect application for a router? The router is in a perfect position to verify that I am on its network.

OK ... how?

> Yes, as evident from the absolute lack of overlapping authentication and encryption standards...

... implemented in the same product, where one of them would always have been enough to meet the requirements of 99% of potential users, and the others would have taken considerably more effort to implement?

> Not necessarily, but the designer of a standard could take both use cases into account.

Which doesn't help if it's a separate mechanism. If 1% of the work gets you to the goal in 99% of the cases, that's what vendors will do. Whether that fulfills the requirements of some standard or not does not matter.

> What, more precisely?

I am not making any suggestions as to the solution.

> Of course taking into consideration that 100% of users might not even have an internet connection.

So, if the device is one that does not inherently need global internet connectivity to be useful, then, yeah, things should work without global internet connectivity.

> One thing I'd like is for devices to a their key signature printed on a sticker. Then I can verify the signature, log in and generate a new key and password, generate a certificate that I can install myself or sign up with a service like Let's Encrypt.

Well, a fixed key is a problem, but other than that, yeah, an out-of-band path for key exchange sounds good.

> No, it's not misleading to say that they are allocated for private use. Your ISP drops connections to these addresses because they respect RFC1819 and don't route to the private address ranges. Even if they didn't, these address ranges are still allocated for private use, and your ISP is Wrong. They're only routable in the sense that IP would technically allow it, but the internet is not simply IP but a collection of standards and best practices.

OK, let's get this straight: What does "private" mean? It's a word with a whole lot of only partially overlapping definitions. For the purposes of this discussion, it is important to distinguish the aspect of "independent from official entities" from the aspect of "not revealed to the public", i.e. "providing privacy". RFC1918 addresses are only private in the former sense: You can allocate and use them without coordinating with RIRs or your ISP. However, they have absolutely nothing to do with the latter sense of providing privacy. That is why it is misleading to call them "private addresses": People understand that to mean that they are defined to provide some sort of secrecy or privacy or protection from the public or something along those lines, which they don't. It's not wrong, because there is a different meaning of "private" that fits exactly what RFC1918 are defined to be used for, but it is misleading because it leads people to assume that it encompasses more than that.

Also, whether ISPs do it or not doesn't really matter. What matters is that RFC1918 address space is in fact routed between networks that are not intended to trust each other. And that is perfectly within the uses intended in RFC1918. The RFC isn't concerned with home networks, really, but with "enterprises", and it defines "an enterprise" to be the scope of an RFC1918 allocation. Nowhere does it say that that implies any sort of trust or security relationship between machines within such an allocation. And also, in practice, it is common to link RFC1918 networks of different "enterprises" together, as a sort-of "meta-enterprise", where a trust relationship is even less likely.

The only thing that is "private" about RFC1918 addresses is that you can allocate them without coordination with IANA/RIRs/ISPs, and that you cannot expect an ISP to route them for you on the global internet. There is no privacy specified in the RFC.

> And sure, "private" has a very broad meaning. A browser could very well flag a certificate that was distributed from a private address as such and let the user decide whether they trust that source.

How does it matter for this whether the address is "private" (i.e., allocated without coordination with IANA/RIRs/ISPs)?

> Sure. But again, "in practice means when you build an IPv6 network", i.e. not a typical consumer, for how long more?

Erm, most IPv6 use is by consumers, with ~ 20% adoption based on google users?! Not sure whether that's quite "typical" yet, but certainly not unusual either. Most user devices support IPv6, and increasingly, ISPs are rolling out IPv6 to their customers with new subscriptions, which tends to come with new routers, which means that at that point their network is using IPv6 for all services that support it.

> Well, it involves IPv6, we can start there. We're talking about a new security policy that a major browser seems to want to implement shortly, definitely much more shortly than full IPv6 rollout.

Yes, and that is the only way to do it. If you wait until after full IPv6 rollout, you will have to work around assumptions that device vendors by then will have made based on the browser's behaviour, which means it only gets harder to implement. If you want to have any hope of success, you have to act now, when your actions can shape what device vendors will do.

> This is a very loaded question, given that we still disagree on whether a solution that works well both for globally routable and NATed devices is possible.

You have so far failed to even show a solution that works better for NATed devices than non-NATed ones.

> I never said that no one uses IPv6, so I agree that you don't follow.

Replace "noone" with "essentially noone" if you want to get my point.

> Let's say that I see your ladder. It's broken, so I tell you that it's unsafe. Unreasonable assumption? You take it down and bring another ladder. I don't see it, but I tell you it's unsafe. You see it and can clearly say that it isn't. Is it unsafe? Is it reasonable for me to tell you that it is unsafe?

Unsafety is not a(n objective) property of the ladder, it's a (subjective) state of your knowledge. The ladder will only either fail or not (that is an objective fact about the ladder). Even a ladder with partially broken steps might still hold up, and a ladder that is all new and shiny can still have some manufacturing defect that causes it to fail on first use. The former is good to use, the latter is not. But that is a useless concept if your goal is to minimize harm because you only know that after the fact. So, what we use instead is a concept of "unsafety". Statements about unsafety are an expression of our knowledge about something. So, the ladder with broken steps is considered unsafe, because based on what we generally know about the statistical properties of ladders with broken steps, they are known to have an increased failure rate. But then, you might apply load tests to that ladder and establish that it does carry the loads required reliably if you avoid the obviously broken steps, in which case it can be considered not unsafe. Mind you, nothing has changed about the ladder, only our knowledge about it has changed. Similarly for the new and shiny ladder, those are generally considered not unsafe because of what we statistically know about new and shiny ladders, and maybe about how ladders are tested after manufacturing. But then, you could test that as well, and maybe find that it breaks apart under light load, at which point you would change to considering it unsafe. Again, nothing has changed about the ladder, it's all about the knowledge you have about it. And the tests I suggested are not the end of that process of discovering the unsafety of a thing. You might still do other tests yet and come to yet another conclusion (like, I dunno, the testing conditions were unnecessarily harsh, and under more realistic usage conditions the opposite conclusion is appropriate).

Now, not knowing anything about the ladder is just another state of knowledge. And if your goal is to minimize harm, then the default is not to assume safety. Again, that is in no way a statement about the ladder. That does not mean the ladder won't hold up. That only means that the ladder is not known (to you!) to hold up. It is always and exclusively a statement about your knowledge about the ladder.

This is not about answering the question "will the ladder fail?", this is about answering the question "is it known to the best of our understanding that the ladder will not fail under some generally expected load conditions?". If the answer to that is "no", then that is reason to be cautious, and that is why the browser warns you/is going to warn you.

You can of course argue that your goal is not to minimize harm, in which case the default assumption does not apply ... but then the whole discussion is pointless, as you are then essentially just saying "if you don't care about minimizing harm, there is no problem with trusting unencrypted connections (of some sort or another)". True, but not my goal, and obviously also not the goal of those people implementing the change.

> Maybe that's actually the better option. But no, "security unknown" in that sense is not equivalent to insecure. As an extreme, I could create a network with an Ethernet cable between two off-grid devices that I control in a faraday cage.

Yes, you could. But the browser doesn't know that. Therefore, its subjective determination is "this is not known to me to protect your private data", and that is what it is telling you. If you know better, that's fine, but the browser doesn't, so it warns you. If you don't know better, you better should listen to what your browser is telling you if your goal is to minimize harm. If you do know better, why do you care that your browser warns you based on its incomplete knowledge about the world?

B) Scopes are exactly: "the global internet and the home".

Considering those things why should it be absurd if I want to secure my home scope at the application layer too? IPv6 is literally designed to allow for this. Browsers are the things that are being stubborn.

If IPv6 is just a terrible collection of hacks then we need a new version and fast before everyone get stuck on v6 for the next 50 years....

1. global addresses,

2. rotating temporary global addresses,

3. universal local addresses (one for each site), and

4. a link-local address (required).

The first 3 now technically reside in the global scope. ULAs used to be called site-local and had their own scope, but they were restructured to basically be fancy UUIDs and their scope abolished and merged with the global scope. Link-local is still its own scope.

Although both are globally scoped, there's a difference between a global IPv6 address and a ULA. The global is globally routeable and prefixes are organized regionally and delegated to allow hierarchical routing while ULAs have arbitrary prefixes (not suitable for global routing) and are not supposed to be forwarded to interfaces outside their subnet.

So to answer your question, for local communications in your home that you didn't want leaving your network, you would use ULAs. You could use link-local addresses if your home was all on the same l2 link, but the generally preferred solution is to use ULAs so you don't leak protocol details upwards and so you can leverage l3 tunnels.

Local DNS is allowed to respond with ULAs, just not servers participating in the global authoritative DNS. If you want DNS on your home site you simply run a local DNS server that resolves your local names and is configured to forward unknown names to the global DNS.

IPv6 kills NAT, so "scoped" addresses step in to fill the void and are overall a much better solution.

Well, not disagreeing that you can have those, but why are you mentioning that specifically for "global addresses", but not for ULA?

> while ULAs have arbitrary prefixes (not suitable for global routing)

How are ULA prefixes different from PI allocations as far as suitability for global routing is concerned?

> and are not supposed to be forwarded to interfaces outside their subnet.

Would you mind pointing to the normative source for this forwarding policy?

> Local DNS is allowed to respond with ULAs, just not servers participating in the global authoritative DNS.

Would you mind pointing to the normative source for this?

> IPv6 kills NAT, so "scoped" addresses step in to fill the void and are overall a much better solution.

But you said that ULA are global scope (which they are, I agree), so what does this have to do with scoped addresses?

Also, which void exactly does killing NAT leave that needs filling (and that then supposedly is filled by ULA)?