March 29, 2024

Web Certification Fail: Bad Assumptions Lead to Bad Technology

It should be abundantly clear, from two recent posts here, that the current model for certifying the identity of web sites is deeply flawed. When you connect to a web site, and your browser displays an https URL and a happy lock or key icon indicating a secure connection, the odds that you’re connecting to an impostor site, despite your browser’s best efforts, are uncomfortably high.

How did this happen? The last two posts unpacked some of the detailed problems with the current system. Today I want to explore the root cause: today’s system is based on wildly unrealistic assumptions about organizations and trust.

The theory behind the system is simple. Browser vendors will identify a set of Certificate Authorities (CAs) who are trusted to certify identities. Browsers will automatically accept any identity certificate issued by any of the trusted CAs.

The first step in making this system work is identifying some CA who is trusted by everybody in the world.

If that last sentence didn’t strike you as odd, go back and read it again. That’s right, the system assumes that there is some party who is trusted by everyone in the world — a spectacularly naive assumption.

Network engineers like to joke about the “evil bit”, a hypothetical label put on each network packet, indicating whether the packet is evil. (See RFC 3514, Steve Bellovin’s classic parody standards document codifying the evil bit. I’ve always loved that the official Internet standards series accepts parody standards.) Well, the “trusted bit” for certificate authorities is pretty much as the same as the evil bit, only applied to organizations rather than network packets. Yet somehow we ended up with a design that relies on this “trusted bit”.

The reason, in part, is unclear thinking about institutional trust, abetted by the unclear language we often use in discussing trust online. For example, we tend to conflate two meanings of the word “trusted”. The first meaning of “trusted”, which is the everyday meaning, implies a judgment that a party is unlikely to misbehave. The second meaning of “trusted”, more common in military security settings, is a factual statement that someone is vulnerable to misbehavior by another. In an ideal world, we would make sure that someone was trusted in the first sense before they became trusted in the second sense, that is, we would make sure that someone was unlikely to misbehave before we we made ourselves vulnerable to their misbehavior. This isn’t easy to do — and we will forget entirely to do it if we confuse the two meanings of trusted.

The second linguistic problem is to use the passive-voice construction “A is trusted to do X” rather than the active form “B trusts A to do X.” The first form is problematic because it doesn’t say who is doing the trusting. Consider these two statements: (A) “CNNIC is a trusted certificate authority.” (B) “Everyone trusts CNNIC to be a certificate authority.” The first statement might sound plausible, but the second is obviously false.

If you try to explain to yourself why the existing web certification system is sound, while avoiding the two errors above (confusing two senses of “trusted”, and failing to say who is doing the trusting), you’ll see pretty quickly that the argument for the current system is tenuous at best. You’ll see, too, that we can’t fix the system by using different cryptography — what we need are new institutional arrangements.