Articles / Cryptography

  • cryptography
  • key-escrow
  • encryption
  • clipper-chip
  • lawful-intercept
  • security

Why a Backdoor Only the Good Guys Can Open Fails as Key Escrow

The pitch is always the same: build a lock with two keys, one for the user and one held by the authorities, and only the right people will ever turn the second. It sounds like a normal safe with a spare key in the bank. It fails for reasons that are mostly engineering, not politics, and we have thirty years of real examples.

Nothing in the maths knows who is holding the key. A key that opens a message for a warranted officer opens it for anyone else who gets hold of that key, or of the system that guards it.

The Clipper chip: escrow done by experts, and still broken

In 1993 the US government proposed the Clipper chip. Each chip had a unique unit key, split into two halves and held by two separate escrow agents. Every call also carried a Law Enforcement Access Field (LEAF), which let an agent with a warrant recover the session key.

This was designed by professionals, with split custody and a proper policy process. It still had a protocol flaw. The LEAF contained a 16-bit checksum, and Matt Blaze showed in 1994 that you could forge a LEAF that passed the check without containing a usable escrowed key. The receiving chip accepted it; the wiretapper learned nothing.

Sixteen bits is only 65,536 possibilities. Here is a toy version of the attack, simplified (the real LEAF format and the Skipjack internals are different, I have used an HMAC as a stand-in):

package main

import (
	"crypto/hmac"
	"crypto/rand"
	"crypto/sha256"
	"fmt"
	"log"
)

// check stands in for the receiving chip: it accepts a LEAF body
// if the 16-bit tag matches the one derived from the session key.
func check(sessionKey, body []byte, tag [2]byte) bool {
	m := hmac.New(sha256.New, sessionKey)
	m.Write(body)
	s := m.Sum(nil)
	return s[0] == tag[0] && s[1] == tag[1]
}

func main() {
	key := make([]byte, 16)
	if _, err := rand.Read(key); err != nil {
		log.Fatal(err)
	}
	// The forger wants a body that is NOT a valid escrow blob
	// but still passes the check; here the tag is fixed at 0x0000.
	body := make([]byte, 16)
	for tries := 1; ; tries++ {
		if _, err := rand.Read(body); err != nil {
			log.Fatal(err)
		}
		if check(key, body, [2]byte{0, 0}) {
			fmt.Println("accepted after", tries, "tries")
			return
		}
	}
}

It finishes in tens of thousands of attempts, which is nothing. Blaze's real attack took around 42 minutes on the hardware of the day, using a Fortezza card as the checking oracle.

The three places an escrow scheme breaks

Clipper's bug is one example of a broader pattern. Escrow schemes tend to fail in one of three places:

  • The protocol: the extra machinery for recovery is new code with new bugs, like that checksum.
  • The custody: someone has to hold the key, and that someone is a target.
  • The scope: the key works for whoever can present it, which is not the same as whoever should.

Each is worth a closer look, because "just be more careful" does not fix any of them.

Custody turns many small targets into one big one

Without escrow, an attacker who wants your messages has to compromise you, or your device. With escrow, there is a database (or an HSM, or a ceremony) that can open everybody's traffic. That makes it the single most valuable thing in the system.

Splitting the key helps against one bad insider, which is why Clipper used two agents. It does not help once the operational reality sets in: requests arrive around the clock, in volume, from many agencies and countries. A system built to answer them quickly is a system with many people and many network paths that can reach the key.

Quick detour: the Juniper episode

Dual_EC_DRBG is an NSA-influenced random number generator whose constants could, in principle, hide a trapdoor for whoever generated them. In 2015 Juniper disclosed unauthorised code in ScreenOS. The attackers had not needed to invent a backdoor: they changed the constant used by Dual_EC so that they held the matching secret instead.

Hang on, that is worth sitting with. The mechanism was already there. Someone else just swapped the key. A backdoor is a capability, and capabilities get repurposed. The only people who can say "this one is only for us" are the ones who never get hacked.

Lawful intercept is escrow's real-world test, and it leaks

We do not have to theorise about what happens when a system has a built-in access path for the authorities. Telecoms already have one: lawful intercept interfaces. The record is not reassuring:

  • In the Greek Vodafone affair (2004 to 2005), the intercept features in the network switches were activated without authorisation and used to listen to officials' phones.
  • In 2024, reports described Chinese state-linked intruders, tracked as Salt Typhoon, inside US telecom networks, including systems used for lawful intercept.

I would treat the 2024 reporting as what it is: press and government statements, with details still emerging. The pattern is the point. The access path existed for the right reasons and was used by the wrong people.

It collides with forward secrecy

Modern protocols are built so that old traffic stays safe even if a long-term key leaks later. TLS 1.3 and the Signal protocol use ephemeral keys that are deleted after use. That is a deliberate property, and I have written about it elsewhere on this blog.

Escrow wants the opposite: a recoverable key for every session. To support it you either keep the session keys (destroying forward secrecy), or you wrap each one for an escrow holder (adding a fresh target). There is no third option where the key is both gone and recoverable.

Scope: who counts as the good guys?

The last problem is not technical at all. Whoever you build the access for, the capability does not stay in one jurisdiction. If a UK service provider builds a recovery path for UK warrants, other governments will ask for the same. Some of them have worse oversight than others, and the provider has to decide, in public, which ones qualify.

On the UK law side, the Investigatory Powers Act 2016 allows the Home Secretary to issue technical capability notices to telecoms operators. How far those can reach into end-to-end encryption is contested, and it depends on the particular notice and operator, so treat any firm claim either way with suspicion, mine included.

What the 2015 paper said

In 2015 a group of well-known cryptographers and security researchers published "Keys Under Doormats", arguing that exceptional access would add complexity, break forward secrecy and create concentrated targets. Their argument was not that it is impossible to build a lock with a second key. It was that you cannot build one that is as secure as the lock without it, at internet scale, and keep it secure over decades.

That is the honest version of the objection. Escrow is possible. A system that is escrowed and also as safe as one that is not has not been shown to exist, and the examples above are what the attempts looked like.