How do you know if an RNG is working?

No matter how much cryptographers accomplish, we’re always
building on a questionable foundation. (illustration: Marc S. Rousseau)

Last week, Edward Snowden spoke to a packed crowd at SXSW about the many problems (and limited solutions) facing those of us who want to keep our communications private. Snowden said a number of things — including a shout out to Moxie’s company Whisper Systems, who certainly deserve it. But instead of talking about that, I wanted to focus on (in my opinion) one of Snowden’s most important quotes:

We need all those brilliant Belgian cryptographers to go “alright we know that these encryption algorithms we are using today work, typically it is the random number generators that are attacked as opposed to the encryption algorithms themselves. How can we make them [secure], how can we test them?”

Now it’s possible I’m a little biased, but it seems to me this cuts to the core of our problems with building secure systems in an increasingly hostile world. Namely: most encryption relies on some source of “random” numbers, either to generate keys or (particularly in the case of public key encryption) to provide semantic security for our ciphertexts.

What this means is that an attacker who can predict the output of your RNG — perhaps by taking advantage of a bug, or even compromising it at a design level — can often completely decrypt your communications. The Debian project learned this firsthand, as have many others. This certainly hasn’t escaped NSA’s notice, if the allegations regarding its Dual EC random number generator are true.

All of this bring us back to Snowden’s quote above, and the question he throws open for us. How do you know that an RNG is working? What kind of tests can we run on our code to avoid flaws ranging from the idiotic to the highly malicious? Unfortunately this question does not have an easy answer. In the rest of this post I’m going to try to explain why.

Background: Random and Pseudorandom Number Generation

I’ve written quite a bit about random number generation on this blog, but before we go forward it’s worth summarizing a few basic facts about random number generation.

First off, the ‘random’ numbers we use in most deployed cryptographic systems actually come from two different systems:

  1. A ‘true’ random number generator (or entropy generator) that collects entropy from the physical world. This can include entropy collected from low-level physical effects like thermal noise and shot noise, or it can include goofy stuff like mouse movements and hard disk seek times.
  2. An algorithmic ‘pseudorandom number generator’ (PRNG) that typically processes the output of (1) to both stretch the output to provide more bits and, in some cases, provide additional security protections in case the output of (1) proves to be biased.
It’s important to note that pseudorandom number generators aren’t “random number generators” at all. These generators typically use cryptographic algorithms (e.g., block ciphers or hash functions) to process a seed value from the RNG into many apparently random looking and unpredictable bytes.

In most cases, it’s quite rare for your application to ever see the raw output of a true random number generator.* Even the low-level entropy collector within Linux’s RNG uses cryptographic constructs like hash functions in order to ‘mix’ the output of various entropy sources. To produce the bits produced in /dev/random or /dev/urandom Linux then seeds a PRNG like Yarrow or Fortuna.**

Another similar pattern occurs inside of the Intel “secure key” random number generator included in Intel Ivy Bridge processors. When you buy one of these processors, you’re getting (free!) a hardware ‘1-shot‘ circuit that collects high-entropy electronic noise, which is then measured and processed into useful RNG output. The design looks like this:

Hardware random number generator used on Intel Ivy Bridge processors. Left: the ‘1-shot’ circuit used to collect physical entropy. Right: the data flow from generator to output, including health checks and
PRNG computation. (source).

Once again, with Intel’s design you (i.e., the application developer) don’t get access to this raw randomness. It’s first used to seed a PRNG based on AES (CTR-DRBG from NIST SP800-90A). What you actually get as an application developer is the processed output of that algorithm.

In practice this typical design some implications. On the positive side, the presence of a PRNG means that the underlying RNG circuit can get pretty borked (e.g., biased) without the results being detectable by your application. On the negative side, the underlying RNG circuit can get pretty borked without the results being detectable in your application.

In other words, with only a few ugly glitches — things that can happen in real life — you can easily get a broken random number generator that nobody notices until it’s way too late. And that’s without deliberate tampering, which makes things way, way worse.

Which brings us back to our fundamental question: how do systems know that their RNG is working. This turns out to be a question without a perfect answer.

Statistical Tests

If you look at the literature on random number generators, you’ll find a lot of references to statistical randomness testing suites like Diehard or NIST’s SP 800-22. The gist of these systems is that they look a the output of an RNG and run tests to determine whether the output is, from a statistical perspective, “good enough” for government work (very literally, in the case of the NIST suite.)

The nature of these tests varies. Some look at simple factors like bias (the number of 1s and 0s) while others look for more sophisticated features such as the distribution of numbers when mapped into 3-D space.

Now I don’t want to knock these tests. They’re a perfectly valid way to detect serious flaws in a (true) RNG — I can attest to this, since I’ve built one that failed the tests miserably — but they probably won’t detect flaws in your system. That’s because like I said above, most deployed systems include a combination of RNG and PRNG, or even RNG plus “conditioning” via cryptographic hash functions or ciphers. The nature of these cryptographic, algorithmic processes is such that virtually every processed output will pass statistical tests with flying colors — even if the PRNG is initialized with ‘garbage’ input.

This means, unfortunately, that it can be very hard to use statistical tests to detect a broken RNG unless you properly test it only at the low level. And even there you won’t rule out intentional backdoors — as I’ll discuss in a moment.

Known Answer Tests (KATs)

Assuming that you’ve tested your true RNG properly and it’s passing all tests, it’s still important to test your PRNG. One approach to doing this is to use Known Answer Tests (KATs) that are essentially test vectors. These contain some input seed material as well as a set of output bytes that should be the algorithmic result of running the PRNG on that seed.

Since PRNGs are purely algorithmic, the theory here is that you can test them like algorithms. While this approach is valid, it raises two potential issues (both of which I’ve seen in practice).

First, you can only test your PRNG on so many points. Thus it’s quite possible that your PRNG will succeed on one particular test vector (i.e., it’ll output just so many valid bytes) but go completely off the reservation on some other input. This is unlikely, but not impossible in normal conditions. It’s very possible if someone is trying to build a malicious backdoor into your PRNG implementation.

Second, the process of instrumenting your PRNG implementation for testing can actually introduce vulnerabilities in your deployed system! Think about this for a second. Normal PRNGs take in real random seeds from your RNG. The last thing you’d ever want to do is run your PRNG on some predictable seed — if you did, everyone would be able to predict the PRNGs outputs. Yet adding a test harness your system means building in logic to re-seed your RNG to something predictable!

This is like adding an ejection seat to your car. Might make you safer — unless it goes off while you’re driving to work.

A quick glance through e.g., the OpenSSL code shows that indeed, exactly this sort of code exists and ships in some versions of the library. Of course, the experienced developers will note that surely such features could be surrounded by pre-processor directives (or the equivalent in your language of choice) ensuring that they’ll never be activated in production code. Sadly, at least in the case of FIPS, this is not possible — for reasons I’ll explain next.

Runtime Health Checks

Another approach to testing RNGs is to test them while the system is running. This isn’t intended to rule out design-level flaws (as the above statistical and KAT tests are) but it is intended to catch situations where the RNG becomes broken during normal operation. This can occur for a variety of reasons, e.g., manufacturing defects, system damage, and even exposure to outside radiation.

Health checks can take different forms. FIPS 140, for example, mandates that all approved RNGs be tested at startup time using KATs. (This is why you can’t make your test harness conditional on compilation flags — it must ship in your production code!) They subsequently mandate a runtime health check that verifies the generator has not become ‘stuck’, i.e., is spitting out the same bytes over and over again.

While I’m sure this last test may have saved someone, somewhere, it seems totally inappropriate and useless when applied to the output of an RNG/PRNG pair, which is how NIST recommends it be used. This is because even the most broken algorithmic PRNGs will almost never spit out duplicate values — even if the underlying RNG fails completely.

The upshot of this decision is that NIST (FIPS) recommend a check that will almost never succeed in catching anything useful from a PRNG, but does introduce a whole bunch of extra logic that can suffer from flaws and/or malicious circumvention. I’m sure the good folks at NIST realize this, but they recommend it anyway — after all, what else are they going to do?


Which brings us to the $10 million question. What happens if an attacker is deliberately tampering with our RNG/PRNG in order to make it fail? Note that this is not an academic question. We have excellent reason to believe it’s happened in some real systems.

Just for fun, let’s go back to the Intel Ivy Bridge RNG described above. We’ll take a look specifically at the PRNG portion of the design, which uses the NIST CTR-DRBG random number generator with AES:

Portion of the Intel Ivy Bridge design, with a few annotations added by yours truly. (original source)

The CTR-DRBG design relies on two features. First, an AES key is selected at random along with some input seed. This pair goes into the AES cipher, where it is processed to derive a new key and data. The result should be unpredictable to most attackers.

But if you were able to change the way keys were updated (in the key_in_mux hilighted) so that instead of updating the key and/or using an unpredictable one, it chose a fixed key known to the attacker, you would now have a very powerful backdoor. Specifically, the output would still look statistically perfectly random. But an attacker who knows this key could simply decrypt one block of RNG output to obtain all future and past outputs of the generator until the next time it was reseeded.

Note that I am not saying the Intel system has a backdoor in it — far from it. I’m only considering how easily it might be made to have one if you were an attacker with control of Intel’s fabrication plants (or their microcode updates). And this is hardly Intel’s fault. It’s just the nature of this particular RNG design. Others could be just as vulnerable.

Actually using this knowledge to attack applications would be more complex, since many system-level RNGs (including the Linux Kernel RNG) combine the output of the RNG with other system entropy (through XOR, unfortunately, not hashing). But Intel has pushed hard to see their RNG output used directly, and there exist plugins for OpenSSL that allow you to use it similarly. If you used such a method, these hypothetical flaws could easily make their way all the way into your cryptography.

Designing against these issues

Unfortunately, so far all I’ve done is call out the challenges with building trustworthy RNGs. And there’s a reason for this: the challenges are easy to identify, while the solutions themselves are hard. And unfortunately at this time, they’re quite manual.

Building secure RNG/PRNGs still requires a combination of design expertise, careful low-level (true) RNG testing — using expert design and statistical tests — and the use of certified algorithms with proper tests. All of the techniques above contribute to building a secure RNG, but none of them are quite sufficient.

Solving this problem, at least in software, so we can ensure that code is correct and does not contain hidden ‘easter eggs’, represents one of the more significant research challenges facing those of us who depend on secure cryptographic primitives. I do hope some enterprising graduate students will give these issues the attention they deserve.


* Though there are some exceptions. See, for example, this FIPS certified smart card that included a bad RNG which was used to generate cryptographic secrets. In general FIPS disallows this except for a very small number of approved RNGs. Perhaps this was one.

** The original version of this post claimed that /dev/random seeds /dev/urandom. This is actually a mistake — both /dev/random and /dev/urandom use the same PRNG, but /dev/random simply keeps track of how much ‘entropy’ is in the pool and blocks when you have drawn too many bits. Thanks to Brendan Long and Thomas Ptacek for setting me straight.

A few more notes on NSA random number generators

Last Friday, Joseph Menn from Reuters published an article claiming that RSA, the pioneering security firm and division of EMC, accepted $10 million dollars to include the 3a7ac-thomas_adversaryDual EC random number generator as the default in their flagship BSAFE library. I’ve written a bit about Dual EC on this blog, so readers will know that I don’t think highly of it. For one thing, it’s a rotten choice for a commercial product, due to its lousy software performance. Much worse: it may introduce a practical backdoor into any system that uses it.

Given the numerous problems with Dual EC it’s baffling that RSA would select this particular generator as the default for their software. It’s particularly strange given that BSAFE is designed for use in Java-based and embedded systems where performance truly is at a premium. And none of this can be explained by the needs of a single RSA customer, since those needs could be satisfied merely by making BSAFE an option, rather than the default.

Of course there have been many people who had viewed RSA’s decision as potentially malicious. Unsupported rumors have been floating around since long before Reuters got involved. What’s new this time is that RSA insiders appear to be going on the record.

I suppose time will tell how the industry reacts to this news. In the interim I have a few small facts to add to the discussion, so I thought I’d sketch them out in this post.

#1: Dual_EC_DRBG’s ‘backdoor’ was known as of January 2005

It’s widely believed that the ‘vulnerability’ in Dual_EC was first identified by Microsoft employees Dan Shumow and Niels Ferguson in the summer of 2007. Tanja Lange (and nymble) recently tipped me off to the fact that this isn’t precisely true.

In point of fact, the possibility of a backdoor was known to at least some members of the ANSI X9.82 standardization committee as far back in January 2005. This surprising news comes via a patent application filed by Certicom employees Dan Brown and Scott Vanstone. The application claims a priority date of January 2005. Here’s the scary bit:

If P and Q are established in a security domain controlled by an administrator, and the entity who generates Q for the domain does so with knowledge of e (or indirectly via knowledge of d). The administrator will have an escrow key for every ECRNG that follows that standard. 

Escrow keys are known to have advantages in some contexts. They can provide a backup functionality. If a cryptographic key is lost, then data encrypted under that key is also lost. However, encryption keys are generally the output of random number generators. Therefore, if the ECRNG is used to generate the encryption key K, then it may be possible that the escrow key e can be used to recover the encryption key K. Escrow keys can provide other functionality, such as for use in a wiretap. In this case, trusted law enforcement agents may need to decrypt encrypted traffic of criminals, and to do this they may want to be able to use an escrow key to recover an encryption key.

For example, in the SSL and TLS protocols, which are used for securing web (HTTP) traffic, a client and server perform a handshake in which their first actions are to exchange random values sent in the clear.

The patent also describes a number of ways to close the backdoor in Dual_EC_DRBG. Indeed, it may be due to Brown and Vanstone that the NIST standard includes an alternative method to close the backdoor (by generating a random Q point).

The existence of this patent does not mean that Brown and Vanstone were responsible for Dual EC. In fact, the generator appears to be an NSA invention, and may date back to the early 2000s. What this patent demonstrates is that some members of the ANSI committee, of which RSA was also a member, had reason to at least suspect that Dual EC could be used to create a wiretapping backdoor. (Update: John Kelsey confirms this.) It would be curious to know how widely this information was shared, and whether anyone on the committee ever inquired as to the provenance of the default parameters.

#2. Dual_EC_DRBG is not really a NIST standard

This is hardly a secret, but it’s something that hasn’t been widely acknowledged in the press. Dual_EC_DRBG is generally viewed as a NIST standard, since it was published in NIST Special Publication 800-90. But that’s not the only place it appears, nor was it developed at NIST.

A complete history of Dual_EC_DRBG would begin with NSA’s drive to include it in the ANSI X9.82 DRBG standard, with a standardization process kicked off in the early 2000s. The draft ANSI standard includes Dual_EC_DRBG with all of the known parameters, along with several additional elliptic curve parameters that were not included in the NIST standards.

Members of the ANSI X9F1 Tool Standards and Guidelines Group which wrote ANSI X9.82.

However you’ll also find Dual_EC_DRBG in the international standard ISO 18031. In other words, Dual EC was widely propagated in numerous standards long before NIST finalized it, and it’s hardly limited to US borders. The ANSI and ISO drafts are in some sense worse, since they don’t include a technique for generating your own Q parameter.

#3. Dual_EC_DRBG is not the only asymmetric random number generator in the ANSI and ISO standards

Cryptographers generally think of Dual EC as the only ‘public key’ random number generator to be widely standardized. We also point to NSA’s generation of the public parameters as evidence that Dual_EC may be compromised.

But in fact, the ANSI X9.82 and ISO standards each include a second generator based on public key cryptographic techniques. And like Dual EC, this one ships with a complete set of default parameters! The additional generator is based on an algorithm due to Micali and Schnorr, and relies for its security on assumptions related to the hardness of factoring large composite numbers. It requires an RSA-type modulus, several of which are conveniently provided in the specification.

Two default MS-DRBG moduli from the ISO 18031 specification.

There’s no reason to suspect that MS-DRBG is used by any real products, let alone that there’s a backdoor in the standard. In fact, there’s a curious fact about it: it’s not obvious from the public literature how one would attack the generator even if one knew the factorization of the n values above — though it seems intuitively likely that an attack does exist. Solving the problem, and finding a practical attack for known factorization, would be a fun project for an enthusiastic mathematician.

Since MS-DRBG comes from the same people who brought you Dual EC, if you are using it you might want to think twice.

RSA warns developers not to use RSA products

In today’s news of the weird, RSA (a division of EMC) has recommended that developers desist from using the (allegedly) ‘backdoored’ Dual_EC_DRBG random number generator — which happens to be the default in RSA’s BSafe cryptographic toolkit. Youch.

In case you’re missing the story here, Dual_EC_DRBG (which I wrote about yesterday) is the random number generator voted most likely to be backdoored by the NSA. The story here is that — despite many valid concerns about this generator — RSA went ahead and made it the default generator used for all cryptography in its flagship cryptography library. The implications for RSA and RSA-based products are staggering. In the worst case a modestly bad but by no means worst case, the NSA may be able to intercept SSL/TLS connections made by products implemented with BSafe.

So why would RSA pick Dual_EC as the default? You got me. Not only is Dual_EC hilariously slow — which has real performance implications — it was shown to be a just plain bad random number generator all the way back in 2006. By 2007, when Shumow and Ferguson raised the possibility of a backdoor in the specification, no sensible cryptographer would go near the thing.

And the killer is that RSA employs a number of highly distinguished cryptographers! It’s unlikely that they’d all miss the news about Dual_EC.

We can only speculate about the past. But here in the present we get to watch RSA’s CTO Sam Curry publicly defend RSA’s choices. I sort of feel bad for the guy. But let’s make fun of it anyway.

I’ll take his statement line by line (Sam is the boldface):

“Plenty of other crypto functions (PBKDF2, bcrypt, scrypt) will iterate a hash 1000 times specifically to make it slower.”

Password hash functions are built deliberately slow to frustrate dictionary attacks. Making a random number generator slow is just dumb.

At the time, elliptic curves were in vogue

Say what?

and hash-based RNG was under scrutiny.

Nonsense. A single obsolete hash based generator (FIPS 186) was under scrutiny — and fixed. The NIST SP800-90 draft in which Dual_EC appeared ALSO provided three perfectly nice non-backdoored generators: two based on hash functions and one based on AES. BSafe even implements some of them. Sam, this statement is just plain misleading.

The hope was that elliptic curve techniques—based as they are on number theory—would not suffer many of the same weaknesses as other techniques (like the FIPS 186 SHA-1 generator) that were seen as negative

Dual-EC suffers exactly the same sort of weaknesses as FIPS 186. Unlike the alternative generators in NIST SP800-90 it has a significant bias and really should not be used in production systems. RSA certainly had access to this information after the analyses were published in 2006.

and Dual_EC_DRBG was an accepted and publicly scrutinized standard.

And every bit of public scrutiny said the same thing: this thing is broken! Grab your children and run away!

SP800-90 (which defines Dual EC DRBG) requires new features like continuous testing of the output, mandatory re-seeding,

The exact same can be said for the hash-based and AES-based alternative generators you DIDN’T choose from SP800-90.

optional prediction resistance and the ability to configure for different strengths.

So did you take advantage of any of these options as part of the BSafe defaults? Why not? How about the very simple mitigations that NIST added to SP800-90A as a means to remove concerns that the generator might have a backdoor? Anyone?

There’s not too much else to say here. I guess the best way to put it is: this is all part of the process. First you find the disease. Then you see if you can cure it.

The Many Flaws of Dual_EC_DRBG

The Dual_EC_DRBG generator from NIST SP800-90A.

Update 9/19: RSA warns developers not to use the default Dual_EC_DRBG generator in BSAFE. Oh lord.

As a technical follow up to my previous post about the NSA’s war on crypto, I wanted to make a few specific points about standards. In particular I wanted to address the allegation that NSA inserted a backdoor into the Dual-EC pseudorandom number generator.

For those not following the story, Dual-EC is a pseudorandom number generator proposed by NIST for international use back in 2006. Just a few months later, Shumow and Ferguson made cryptographic history by pointing out that there might be an NSA backdoor in the algorithm. This possibility — fairly remarkable for an algorithm of this type — looked bad and smelled worse. If true, it spelled almost certain doom for anyone relying on Dual-EC to keep their system safe from spying eyes.

Now I should point out that much of this is ancient history. What is news today is the recent leak of classified documents that points a very emphatic finger towards Dual_EC, or rather, to an unnamed ‘2006 NIST standard’. The evidence that Dual-EC is this standard has now become so hard to ignore that NIST recently took the unprecedented step of warning implementers to avoid it altogether.

Better late than never.

In this post I’m going to try to explain the curious story of Dual-EC. While I’ll do my best to keep this discussion at a high and non-mathematical level, be forewarned that I’m probably going to fail at least at a couple of points. I you’re not the mood for all that, here’s a short summary:

  • In 2005-2006 NIST and NSA released a pseudorandom number generator based on elliptic curve cryptography. They released this standard — with very little explanation — both in the US and abroad.
  • This RNG has some serious issues with just being a good RNG. The presence of such obvious bugs was mysterious to cryptographers.
  • In 2007 a pair of Microsoft researchers pointed out that these vulnerabilities combined to produce a perfect storm, which — together with some knowledge that only NIST/NSA might have — opened a perfect backdoor into the random number generator itself.
  • This backdoor may allow the NSA to break nearly any cryptographic system that uses it.

If you’re still with me, strap in. Here goes the long version.


For a good summary on the history of Dual-EC-DRBG, see this 2007 post by Bruce Schneier. Here I’ll just give the highlights.

Back in 2004-5, NIST decided to address a longstanding weakness of the FIPS standards, namely, the limited number of approved pseudorandom bit generator algorithms (PRGs, or ‘DRBGs’ in NIST parlance) available to implementers. This was actually a bit of an issue for FIPS developers, since the existing random number generators had some known design weaknesses.*

NIST’s answer to this problem was Special Publication 800-90, parts of which were later wrapped up into the international standard ISO 18031. The NIST pub added four new generators to the FIPS canon. None these algorithms is a true random number generator in the sense that they collect physical entropy. Instead, what they do is process the (short) output of a true random number generator — like the one in Linux — conditioning and stretching this ‘seed’ into a large number of random-looking bits you can use to get things done.** This is particularly important for FIPS-certified cryptographic modules, since the FIPS 140-2 standards typically require you to use a DRBG as a kind of ‘post-processing’ — even when you have a decent hardware generator.

The first three SP800-90 proposals used standard symmetric components like hash functions and block ciphers. Dual_EC_DRBG was the odd one out, since it employed mathematics more that are typically used to construct public-key cryptosystems. This had some immediate consequences for the generator: Dual-EC is slow in a way that its cousins aren’t. Up to a thousand times slower.

Now before you panic about this, the inefficiency of Dual_EC is not necessarily one of its flaws! Indeed, the inclusion of an algebraic generator actually makes a certain amount of sense. The academic literature includes a distinguished history of provably secure PRGs based on on number theoretic assumptions, and it certainly didn’t hurt to consider one such construction for standardization. Most developers would probably use the faster symmetric alternatives, but perhaps a small number would prefer the added confidence of a provably-secure construction.

Unfortunately, here is where NIST ran into their first problem with Dual_EC.

Flaw #1: Dual-EC has no security proof

Let me spell this out as clearly as I can. In the course of proposing this complex and slow new PRG where the only damn reason you’d ever use the thing is for its security reduction, NIST forgot to provide one. This is like selling someone a Mercedes and forgetting to attach the hood ornament.

I’d like to say this fact alone should have damned Dual_EC, but sadly this is par for the course for NIST — which treats security proofs like those cool Japanese cookies you can never find. In other words, a fancy, exotic luxury. Indeed, NIST has a nasty habit of dumping proof-writing work onto outside academics, often after the standard has been finalized and implemented in products everywhere.

So when NIST put forward its first draft of SP800-90 in 2005, academic cryptographers were left to analyze it from scratch. Which, to their great credit, they were quite successful at.

The first thing reviewers noticed is that Dual-EC follows a known design paradigm — it’s a weird variant of an elliptic curve linear congruential generator. However they also noticed that NIST had made some odd rookie mistakes.

Now here we will have to get slightly wonky — though I will keep mathematics to a minimum. (I promise it will all come together in the end!) Constructions like Dual-EC have basically two stages:

1. A stage that generates a series of pseudorandom elliptic curve points. Just like on the graph at right, an elliptic curve point is a pair (x, y) that satisfies an elliptic curve equation. In general, both x and y are elements of a finite field, which for our purposes means they’re just large integers.***

The main operation of the PRNG is to apply mathematical operations to points on the elliptic curve, in order to generate new points that are pseudorandom — i.e., are indistinguishable from random points in some subgroup.

And the good news is that Dual-EC seems to do this first part beautifully! In fact Brown and Gjøsteen even proved that this part of the generator is sound provided that the Decisional Diffie-Hellman problem is hard in the specific elliptic curve subgroup. This is a well studied hardness assumption so we can probably feel pretty confident in this proof.

2. Extract pseudorandom bits from the generated EC points. While the points generated by Dual-EC may be pseudorandom, that doesn’t mean the specific (x, y) integer pairs are random bitstrings. For one thing, ‘x‘ and ‘y‘ are not really bitstrings at all, they’re integers less than some prime number. Most pairs don’t satisfy the curve equation or are not in the right subgroup. Hence you can’t just output the raw x or y values and expect them to make good pseudorandom bits.

Thus the second phase of the generator is to ‘extract’ some (but not all) of the bits from the EC points. Traditional literature designs do all sorts of things here — including hashing the point or dropping up to half of the bits of the x-coordinate. Dual-EC does something much simpler: it grabs the x coordinate, throws away the most significant 16-18 bits, and outputs the rest.

In 2006, first Gjøsteen and later Schoenmakers and Sidorenko took a close look at Dual-EC and independently came up with a surprising result:

Flaw #2: Dual-EC outputs too many bits

Unlike those previous EC PRGs which output anywhere from 2/3 to half of the bits from the x-coordinate, Dual-EC outputs nearly the entire thing.

This is good for efficiency, but unfortunately it also gives Dual-EC a bias. Due to some quirks in the mathematics of the field operations, an attacker can now predict the next bits of Dual-EC output with a fairly small — but non-trivial — success probability, in the range of 0.1%. While this number may seem small to non-cryptographers, it’s basically a hanging offense for a cryptographic random number generator where probability of predicting a future bit should be many orders of magnitude lower.

What’s just plain baffling is that this flaw ever saw the light of day. After all, the specification was developed by bright people at NIST — in collaboration with NSA. Either of those groups should easily have discovered a bug like this, especially since this issue had been previously studied. Indeed, within a few months of public release, two separate groups of academic cryptographers found it, and were able to implement an attack using standard PC equipment.

So in summary, the bias is mysterious and it seems to be very much an ‘own-goal’ on the NSA’s part. Why in the world would they release so much information from each EC point? It’s hard to say, but a bit more investigation reveals some interesting consequences:

Flaw #3: You can guess the original EC point from looking at the output bits

By itself this isn’t really a flaw, but will turn out to be interesting in just a minute.

Since Dual-EC outputs so many bits from the x-coordinate of each point — all but the most significant 16 bits — it’s relatively easy to guess the original source point by simply brute-forcing the missing 16 bits and solving the elliptic curve equation for y. (This is all high-school algebra, I swear!)

While this process probably won’t uniquely identify the original (x, y), it’ll give you a modestly sized list of candidates. Moreover with only 16 missing bits the search can be done quickly even on a desktop computer. Had Dual_EC thrown away more bits of the x-coordinate, this search would not have been feasible at all.

So what does this mean? In general, recovering the EC point shouldn’t actually be a huge problem. In theory it could lead to a weakness — say predicting future outputs — but in a proper design you would still have to solve a discrete logarithm instance for each and every point in order to predict the next bytes output by the generator.

And here is where things get interesting.

Flaw #4: If you know a certain property about the Dual_EC parameters, and can recover an output point, you can predict all subsequent outputs of the generator

Did I tell you this would get interesting in a minute? I totally did.

The next piece of our puzzle was discovered by Microsoft researchers Dan Shumow and Niels Ferguson, and announced at the CRYPTO 2007 rump session. I think this result can best be described via the totally intuitive diagram below. (Don’t worry, I’ll explain it!)

Annotated diagram from Shumow-Ferguson presentation (CRYPTO 2007).
Colorful elements were added by yours truly. Thick green arrows mean ‘this part is
easy to reverse’. Thick red arrows should mean the opposite. Unless you’re the NSA.

The Dual-EC generator consists of two stages: a portion that generates the output bits (right) and a part that updates the internal state (left).

Starting from the “r_i” value (circled, center) and heading right, the bit generation part first computes the output point using the function “r_i * Q” — where Q is an elliptic curve point defined in the parameters — then truncates 16 bits its off its x-coordinate to get the raw generator output. The “*” operator here describes elliptic point multiplication, which is a complex operation that should be relatively hard to invert.

Note that everything after the point multiplication should be easy to invert and recover from the output, as we discussed in the previous section.

Every time the generator produces one block of output bits, it also updates its internal state. This is designed to prevent attacks where someone compromises the internal values of a working generator, then uses this value to wind the generator backwards and guess past outputs. Starting again from the circled “r_i” value, the generator now heads upwards and computes the point “r_i * P” where P is a different elliptic curve point also described in the parameters. It then does some other stuff.

The theory here is that P and Q should be random points, and thus it should be difficult to find “r_i * P” used for state update even if you know the output point “r_i * Q” — which I stress you do know, because it’s easy to find. Going from one point to the other requires you to know a relationship between P and Q, which you shouldn’t actually know since they’re supposed to be random values. The difficulty of this is indicated by the thick red arrow.

Looks totally kosher to me. (Source: NIST SP800-90A)

There is, however, one tiny little exception to this rule. What if P and Q aren’t entirely random values? What if you chose them yourself specifically so you’d know the mathematical relationship between the two points?

In this case it turns out you can easily compute the next PRG state after recovering a single output point (from 32 bytes of RNG output). This means you can follow the equations through and predict the next output. And the next output after that. And on forever and forever.****

This is a huge deal in the case of SSL/TLS, for example. If I use the Dual-EC PRG to generate the “Client Random” nonce transmitted in the beginning of an SSL connection, then the NSA will be able to predict the “Pre-Master” secret that I’m going to generate during the RSA handshake. Given this information the connection is now a cleartext read. This is not good.

So now you should all be asking the most important question of all: how the hell did the NSA generate the specific P and Q values recommended in Appendix A of Dual-EC-DRBG? And do they know the relationship that allows them to run this attack? All of which brings us to:

Flaw #5Nobody knows where the recommended parameters came from

And if you think that’s problematic, welcome to the club.

But why? And where is Dual-EC used?

The ten million dollar question of Dual-EC is why the NSA would stick such an obviously backdoored algorithm into an important standard. Keep in mind that cryptographers found the major (bias) vulnerabilities almost immediately after Dual-EC shipped. The possibility of a ‘backdoor’ was announced in summer 2007. Who would still use it?

A few people have gone through the list of CMVP-evaluated products and found that the answer is: quite a few people would. Most certify Dual-EC simply because it’s implemented in OpenSSL-FIPS, and they happen to use that library. But at least one provider certifies it exclusively. Yuck.

Hardcoded constants from the OpenSSL-FIPS
implementation of Dual_EC_DRBG. Recognize ’em?

It’s worth keeping in mind that NIST standards carry a lot of weight — even those that might have a backdoor. Folks who aren’t keeping up on the latest crypto results could still innocently use the thing, either by accident or (less innocently) because the government asked them to. Even if they don’t use it, they might include the code in their product — say through the inclusion of OpenSSL-FIPS or MS Crypto API — which means it’s just a function call away from being surreptitiously activated.

Which is why people need to stop including Dual-EC immediately. We have no idea what it’s for, but it needs to go away. Now.

So what about the curves?

The last point I want to make is that the vulnerabilities in Dual-EC have precisely nothing to do with the specifics of the NIST standard elliptic curves themselves. The ‘back door’ in Dual-EC comes exclusively from the relationship between P and Q — the latter of which is published only in the Dual-EC specification. The attack can work even if you don’t use the NIST pseudorandom curves.

However, the revelations about NIST and the NSA certainly make it worth our time to ask whether these curves themselves are somehow weak. The best answer to that question is: we don’t know. Others have observed that NIST’s process for generating the curves leaves a lot to be desired. But including some kind of hypothetical backdoor would be a horrible, horrific idea — one that would almost certainly blow back at us.

You’d think people with common sense would realize this. Unfortunately we can’t count on that anymore.

Thanks to Tanja Lange for her assistance proofing this post. Any errors in the text are entirely mine.


* My recollection of this period is hazy, but prior to SP800-90 the two most common FIPS DRBGs in production were (1) the SHA1-based DSA generator of FIPS 186-2 and (2) ANSI X9.31. The DSA generator was a special-purpose generator based on SHA1, and was really designed just for that purpose. ANSI X9.31 used block ciphers, but suffered from some more subtle weaknesses it retained from the earlier X9.17 generator. These were pointed out by Kelsey, Schneier, Wagner and Hall.

** This is actually a requirement of the FIPS 140-2 specification. Since FIPS does not approve any true random number generators, it instead mandates that you run your true RNG output through a DRBG (PRNG) first. The only exception is if your true RNG has been approved ‘for classified use’.

*** Specifically, x and y are integers in the range 0 to p-1 where p is a large prime number. A point is a pair (x, y) such that y^2 = x^3 + ax + b mod p. The values a and b are defined as part of the curve parameters.

**** The process of predicting future outputs involves a few guesses, since you don’t know the exact output point (and had to guess at the missing 16 bits), but you can easily reduce this to a small set of candidates — then it’s just a question of looking at a few more bits of RNG output until you guess the right one.

Surviving a bad RNG

A couple of weeks ago I wrote a long post about random number generation, which I find to be one of the most fascinating subjects in cryptography — mostly because of how terrible things get when people screw it up.

And oh boy, do people screw it up. Back in 2008 it was Debian, with their ‘custom’ OpenSSL implementation that could only produce 32,768 possible TLS keys (do you really need more?) In 2012 it’s 25,000 factorable TLS public keys, all of which appear to have been generated by embedded devices with crappy RNGs.

When this happens, people get nervous. They start to wonder: am I at risk? And if so, what can I do to protect myself?

Answering this question is easy. Answering it in detail is hard. The easy answer is that if you really believe there’s a problem with your RNG, stop reading this blog and go fix it!

The more complicated answer is that many bad things can happen if your RNG breaks down, and some are harder to deal with than others.

In the rest of this post I’m going to talk about this, and give a few potential mitigations. I want to stress that this post is mostly a thought-exercise. Please do not re-engineer OpenSSL around any of the ‘advice’ I give herein (I’m looking at you, Dan Kaminsky), and if you do follow any of my advice, understand the following:

When it all goes terribly wrong, I’ll quietly take down this post and pretend I never wrote it.

In other words, proceed at your own risk. First, some background.

What’s a ‘bad RNG’?

Before we get started, it’s important to understand what it means for an RNG to be broken. In general, failure comes in three or four different flavors, which may or may not share the same root cause:

  1. Predictable output. This usually happens when a generator is seeded with insufficient entropy. The result is that the attacker can actually predict, or at least guess, the exact bits that the RNG will output. This has all kinds of implications, none of them good.
  2. Resetting output. This can occur when a generator repeatedly outputs the same stream of bits, e.g., every time the system restarts. When an attacker deliberately brings about this condition, we refer to it as a Reset Attack, and it’s a real concern for devices like smartcards.
  3. Shared output. Sometimes exactly the same bits will appear on two or more different devices. Often the owners won’t have any idea this is happening until someone else turns up with their public key! This is almost always caused by some hokey entropy source or hardcoded seed value.

These aren’t necessarily distinct problems, and they can easily bleed into one another. For example, a resetting RNG can become a predictable RNG once the adversary observes the first round of outputs. Shared output can become predictable if the attacker gets his hands on another device of the same model. The Debian bug, for example, could be classified into all three categories.

In addition to the problems above, there’s also a fourth (potential) issue:

4. Non-uniform or biased output. It’s at least possible that the output of your generator will exhibit biased outputs, or strings of repeated characters (the kind of thing that tests like DIEHARD look for). In the worst case, it might just start outputting zero bytes.

The good news is that this is relatively unlikely as long as you’re using a standard crypto library. That’s because modern systems usually process their collected entropy through a pseudo-random generator (PRG) built from a hash function or a block cipher.

The blessing of a PRG is that it will usually give you nice, statistically-uniform output even when you feed it highly non-uniform seed entropy. This helps to prevent attacks (like this one) which rely on the presence of obvious biases in your nonces/keys/IVs. While this isn’t a rule, most common RNG failures seem to be related to bad entropy, not to some surprise failure of the PRG.

Unfortunately, the curse is that a good PRG can hide problems. Since most people only see the output of their PRG (rather than the seed material), it’s easy to believe that your RNG is doing a great job, even when it’s actually quite sick.* This has real-world implications: many standards (like FIPS-140) perform continuous tests on the final RNG/PRG output (e.g., for repeated symbols). The presence of a decent PRG renders these checks largely useless, since they’ll really only detect the most catastrophic (algorithmic) failures.

Key generation

When it comes to generating keys with a bad (P)RNG, the only winning move is not to play. Algorithm aside, if an attacker can predict your ‘randomness’, they can generate the same key themselves. Game over. Incidentally, this goes for ephemeral keys as well, meaning that protocols like Diffie-Hellman are not secure in the presence of a predictable RNG (on either side).

If you think there’s any chance this will happen to you, then either (a) generate your keys on a reliable device, or (b) get yourself a better RNG. If neither option is available, then for god’s sake, collect some entropy from the user before you generate keys. Ask them to tap a ditty on a button, or (if a keyboard is available), get a strong, unpredictable passphrase and hash it through PBKDF2 to get a string of pseudo-random bits. This might not save you, but it’s probably better than the alternative.

What’s fascinating is that some cryptosystems are more vulnerable to bad, or shared randomness than others. The recent batch of factorable RSA keys, for example, appears to be the product of poor entropy on embedded devices. But the keys weren’t broken because someone guessed the entropy that was used. Rather, the mere fact that two different devices share entropy was enough to make both of their keys factorable.

According to Nadia Heninger, this is an artifact of the way that RSA keys are generated. Every RSA public modulus is the product of two primes. Some devices generate one prime, then reseed their RNG (with the time, say) before generating the second. The result is two different moduli, each sharing one prime. Unfortunately, this is the worst thing you can do with an RSA key, since anyone can now compute the GCD and efficiently factor both keys.

Although you’re never going to be safe when two devices share entropy, it’s arguable that you’re better off if they at least generate the same RSA key, rather than two moduli with a single shared prime. One solution is to calculate the second prime as a mathematical function of the first. An even easier fix is just to make sure that you don’t reseed between the two primes.

Of course it’s not really fair to call these ‘solutions’. Either way you’re whistling past the graveyard, but at least this might let you whistle a bit longer.

Digital signatures and MACs

There’s a widely held misconception that digital signatures must be randomized. This isn’t true, but it’s understandable that people might think this, since it’s a common property of the signatures we actually use. Before we talk about this, let me stipulate that what we’re talking about here is the signing operation itself — I’m premising this discussion on the very important assumption that we have properly-generated keys.

MACs. The good news is that virtually every practical MAC in use today is deterministic. While there are probabilistic MACs, they’re rarely used. As long as you’re using a standard primitive like HMAC, that bad RNG shouldn’t affect your ability to authenticate your messages.

Signatures. The situation with signatures is a bit more complicated. I can’t cover all signatures, but let’s at least go over the popular ones. For reasons that have never been adequately explored, these are (in no particular order): ECDSA, DSA, RSA-PKCS#1v1.5 and RSA-PSS. Of these four signatures, three are randomized.

The major exception is RSA-PKCS#1v1.5 signature padding, which has no random fields at all. While this means you can give your RNG a rest, it doesn’t mean that v1.5 padding is good. It’s more accurate to say that the ‘heuristically-secure’ v1.5 padding scheme remains equally bad whether you have a working RNG or not.

If you’re signing with RSA, a much better choice is to use RSA-PSS, since that scheme actually has a reduction to the hardness of the RSA problem. So far so good, but wait a second: doesn’t the P in PSS stand for Probabilistic? And indeed, a close look at the PSS description (below) reveals the presence of random salt in every signature.

The good news is that this salt is only an optimization. It allows the designers to obtain a tighter reduction to the RSA problem, but the security proof holds up even if you repeat the salt, or just hardcode it to zero.

The PSS signing algorithm. MGF is constructed from a hash function.

Having dispensed with RSA, we can get down to the serious offenders: DSA and ECDSA.

The problem in a nutshell is that every (EC)DSA signature includes a random nonce value, which must never be repeated. If you ever forget this warning — i.e., create two signatures (on different messages) using the same nonce — then anyone can recover your secret key. This is both easy to detect, and to compute. You could write a script to troll the Internet for repeated nonces (e.g., in X509 certificates), and then outsource the final calculation to a bright eighth-grader.

Usually when DSA/ECDSA go wrong, it’s because someone simply forgot to generate a random nonce in the first place. This appears to be what happened with the Playstation 3. Obviously, this is stupid and you shouldn’t do it. But no matter how careful your implementation, you’re always going to be vulnerable if your RNG starts spitting out repeated values. If this happens to you even once, you need to throw away your key and generate a new one.

There are basically two ways to protect yourself:

  • Best: don’t to use (EC)DSA in the first place. It’s a stupid algorithm with no reasonable security proof, and as a special bonus it goes completely pear-shaped in the presence of a bad RNG. Unfortunately, it’s also a standard, used in TLS and elsewhere, so you’re stuck with it.
  • Second best: Derive your nonces deterministically from the message and some secret data. If done correctly (big if!), this prevents two messages from being signed with the same nonce. In the extreme case, this approach completely eliminates the need for randomness in (EC)DSA signatures.There are two published proposals that take this approach. The best is Dan Bernstein’s (somewhat complex) EdDSA proposal, which looks like a great replacement for ECDSA. Unfortunately it’s a replacement, not a patch, since EdDSA uses different elliptic curves and is therefore not cross-compatible with existing ECDSA implementations.

    Alternatively, Thomas Pornin has a proposal up that simply modifies (EC)DSA by using HMAC to derive the nonces. The best part about Thomas’s proposal is that it doesn’t break compatibility with existing (EC)DSA implementations. I will caution you, however: while Thomas’s work looks reasonable, his proposal is just a draft (and an expired one to boot). Proceed at your own risk.


There are various consequences to using a bad RNG for encryption, most of which depend on the scheme you’re using. Once again we’ll assume that the keys themselves are properly-generated. What’s at stake is the encryption itself.

Symmetric encryption. The good news is that symmetric encryption can be done securely with no randomness at all, provided that you have a strong encryption key and the ability to keep state between messages.

An obvious choice is to use CTR mode encryption. Since CTR mode IVs needn’t be unpredictable, you can set your initial IV to zero, then simply make sure that you always hang onto the last counter value between messages. Provided that you never ever re-use a counter value with a given key (even across system restarts) you’ll be fine.**

This doesn’t work with CBC mode, since that actually does require an unpredictable IV at the head of each chain. You can hack around this requirement in various ways, but I’m not going to talk about those here; nothing good will come of it.

Public-key encryption. Unfortunately, public-key encryption is much more difficult to get right without a good RNG.

Here’s the fundamental problem: if an attacker knows the randomness you used to produce a ciphertext, then (in the worst case) she can simply encrypt ‘guess’ messages until she obtains the same ciphertext as you. At that point she knows what you encrypted.***

Obviously this attack only works if the attacker can guess the message you encrypted. Hence it’s possible that high-entropy messages (symmetric keys, for example) will encrypt securely even without good randomness. But there’s no guarantee of this. Elgamal, for example, can fail catastrophically when you encrypt two messages with the same random nonce.****

Although I’m not going to endorse any specific public-key encryption scheme, it seems likely that some schemes will hold up better than others. For example, while predictably-randomized RSA-OAEP and RSA-OAEP+ will both be vulnerable to guessing attacks, there’s some (intuitive) reason to believe that they’ll remain secure for high-entropy messages like keys. I can’t prove this, but it seems like a better bet than using Elgamal (clearly broken) or older padding schemes like RSA-PKCS#1v1.5.

If my intuition isn’t satisfying to you, quite a lot of research is still being done in this area. See for example, recent works on deterministic public-key encryption, or hedged public-key encryption. Note that all of this work makes on the assumption that you’re encrypting high-entropy messages.


I can’t conclude this post without at least a token discussion of how a bad RNG can affect cryptographic protocols. The short version is that it depends on the protocol. The shorter version is that it’s almost always bad.

Consider the standard TLS handshake. Both sides use their RNGs to generate nonces. Then the client generates a random ‘Pre-master Secret’ (PMS), encrypts it under the server’s key, and transmits it over the wire. The ‘Master Secret’ (and later, transport key) is derived by hashing together all of the nonces and the PMS.

Since the PMS is the only real ‘secret’ in the protocol (everything else is sent in the clear), predicting it is the same as recovering the transport key. Thus TLS is not safe to use if the client RNG is predictable. What’s interesting is that the protocol is secure (at least, against passive attackers) even if the server’s RNG fails. I can only guess that this was a deliberate choice on the part of TLS’s designers.

SSL handshake (source).

Protocols are already plenty exciting when you have a working RNG. Adding a bad RNG to the mix is like pouring fireworks on a fire. It’s at least possible to build protocols that are resilient to one participant losing their RNG, but it’s very tricky to accomplish — most protocols will fail in unexpected ways.

In Summary

If you take nothing else from this post, I hope it’s this: using a broken RNG is just a bad idea. If you think there’s any chance your generator will stop working, then for god’s sake, fix it! Don’t waste your time doing any of the stuff I mention above.

That said, there legitimately are cases where your RNG can go wrong, or where you don’t have one in the first place. The purpose of this post was to help you understand these scenarios, and the potential consequences for your system. So model it. Think about it. Then spend your time on better things.


 The classic example is Debian’s 2008 OpenSSL release, which used a 15-bit process ID as the only seed for its PRG. This wasn’t obvious during testing, since the 32,768 possible RNG streams all looked pretty random. It was only after the public release that people noticed that many devices were sharing TLS keys.

** If you’re going to do this, you should also be sure to use a MAC on your ciphertext, including the initial counter value for each message.

*** A great example is unpadded, textbook RSA. If m is random, then it’s quite difficult to recover m given m^e mod N. If, however, you have a few good guesses for m and you know the public key (N, e), you can easily try each of your guesses and compare the results.

**** Given two Elgamal ciphertexts on the same key and randomness (g^r, y^r*M1), (g^r, y^r*M2) you can easily compute M1/M2. A similar thing happens with hash Elgamal. This may or may not be useful to you, depending on how much you know about the content of the various messages.

Random number generation: An illustrated primer

Last week we learned (from two different sources!) that certain RSA implementations don’t properly seed their random number generators before generating keys. One practical upshot is that a non-trivial fraction of RSA moduli share a prime factor. Given two such moduli, you can easily factor both.

This key generation kerfuffle is just the tip of the iceberg: a lot of bad things can happen when you use a weak, or improperly-seeded RNG. To name a few:

  • Re-using randomness with (EC)DSA can lead to key recovery.
  • Re-using randomness with Elgamal can lead to plaintext recovery and other ugliness.
  • Using predictable IVs in CBC or CTR mode encryption can lead to plaintext recovery.
  • When protocols use predictable nonces they may become vulnerable to e.g., replay attacks.

In the rest of this post I’m going to talk about the various ways that random number generators work, the difference between RNGs and PRGs, and some of the funny problems with both. Since the post has gotten horrifically long, I’ve decided to present it in a (fun!) question/answer style that makes it easy to read in any order you want. Please feel free to skip around.

What’s the difference between Randomness, Pseudo-Randomness and Entropy?

Before we get started, we have to define a few of our terms. The fact is, there are many, many definitions of randomness. Since for our purposes we’re basically interested in random bit generators, I’m going to give a workaday definition: with a truly random bit generator, nobody (regardless of what information or computing power they have) can predict the next output bit with probability greater than 1/2.

If we lived in an orderly universe, it would be hard to build generators that meet this standard. Fortunately, the universe we live in seems to be anything but orderly. Physicists tell us that at the quantum level certain events have measurable probabilities, but otherwise cannot be predicted in advance.
A hardware RNG.

The most expensive hardware RNGs take advantage of this, measuring such phenomena as radioactive decay or shot noise. Most consumer-grade RNGs don’t have radioactive particles lying around, so they instead measure macroscopic, but chaotic phenomena — typically highly-amplified electrical noise.

These devices are great if you’ve got ’em; unfortunately not everyone does. For the rest of us, the solution is to collect unpredictable values from the computer we’re working on. While this gunk may not be truly random, we hope that it has sufficient entropy — essentially a measure of unpredictability — that our attacker won’t know the difference.

If you’re using a standard PC, your system is probably filling its entropy pool right now: from unpredictable values such as drive seek or inter-keystroke timings. Taken individually none of these events provide enough entropy to do much; but by ‘stirring’ many such measurements together you can obtain enough to do useful cryptography.

Random vs. Pseudorandom. The big problem with RNGs is that they’re usually pretty inefficient. Hardware RNGs can only collect so many bits per second, and the standard OS entropy measurement techniques are even slower. For this reason, many security systems don’t actually use this entropy directly. Instead, they use it to seed a fast cryptographically-secure pseudo-random generator, sometimes called a CSPRNG or (to cryptographers) just a PRG.

PRGs don’t generate random numbers at all. Rather, they’re algorithms that take in a short random string (‘seed’), and stretch it into a long sequence of random-looking bits. Since PRGs are deterministic and computational in nature, they obviously don’t satisfy our definition of randomness (a sufficiently powerful attacker can simply brute-force her way through the seed-space.) But if our attackers are normal (i.e., computationally limited) it’s possible to build unpredictable PRGs from fairly standard assumptions.*

Combining RNGs and PRGs. As I said, most systems combine an RNG with a PRG, using the former to generate a seed for the latter. Some standards actually mandate this combination — not just because it’s faster, but because the additional layer of PRG is believed to offer some resilience in the event that the RNG contains a hardware flaw.

You can argue about whether this is a good idea, but the upshot is as follows: if you want to understand where ‘random’ numbers come from, you really need to understand both technologies and how they interoperate on your machine.

Where does my entropy come from?

Unless you’re running a server and have a fancy Hardware Security Module installed, chances are that your system is collecting entropy from the world around it. Most OSes do this at the kernel level, using a variety of entropy sources which are then ‘stirred’ together. These include:

  • Drive seek timings. Modern hard drives (of the spinning variety) are a wonderful source of chaotic events. In 1994 Davis, Ihaka and Fenstermacher argued that drive seek times are affected by air turbulence within the drive’s enclosure, which makes them an excellent candidate for cryptographic entropy sampling. It’s not clear how this technique holds up against solid-state drives; probably not well.
  • Mouse and keyboard interaction. People are unpredictable. Fortunately for us, that’s a good thing. Many RNGs collect entropy by measuring the time between a user’s keystrokes or mouse movements, then gathering a couple of low-order bits and adding them to the pool.
  • Network events. Although network events (packet timings, for example) seem pretty unpredictable, most systems won’t use this data unless you explicitly tell them to. That’s because the network is generally assumed to be under the adversary’s control (he may be the one sending you those ‘unpredictable’ packets!) You disable these protections at your own risk.
  • Uninitialized memory. Ever forget to initialize a variable? Then you know that RAM is full of junk. While this stuff may not be random, certain systems use it on the theory that it probably can’t hurt. Occasionally it can — though not necessarily in the way you’d think. The classic example is this Debian OpenSSL bug, which (via a comedy of errors) meant that the PRG had only 32,768 possible seed values.
  • Goofy stuff. Some systems will try to collect entropy by conducting unpredictable calculations. One example is to start many threads counting towards infinity, then stop one with a hardware interrupt. I’ve done this once before and evaluated the output. I assure you that YMMV. Significantly.
  • Trusted Platform Module. Many desktop machines these days include a TPM chip on the motherboard. The good news about this is that every TPM contains an internal hardware RNG, which your OS can access if it has the right drivers. It ain’t fast, and the design hasn’t been publicly audited. Still, folding some of this into your entropy pool is probably a good idea.
  • New processor RNGs. To save us all this trouble, the next generation of Intel processors will contain a built-in hardware RNG/PRG, which goes by the codename ‘Bull Mountain’. Perhaps this will be the solution to all of our problems. (h/t David Johnston in comments.)

The upshot of all of this is that on a typical machine there’s usually enough ‘unpredictable’ stuff going on to seed a decent entropy pool. The real problems come up in systems that aren’t typical.

What about VMs and embedded devices?

Life inside an embedded device.

The problem with classical entropy gathering is that it assumes that unpredictable things will actually happen on the system. Unfortunately, VMs and embedded devices defy this expectation, mostly by being very, very boring.

Imagine the following scenario: you have a VM instance running on a server. It has no access to keyboard or mouse input, and only mediated access to hardware, which it shares with eight other VM instances.

Worse yet, your VM may be a clone. Perhaps you just burped up fifty instances of that particular image from a ‘frozen’ state. Each of these VMs may have loads of entropy in its pool, but it’s all the same entropy, across every clone sibling. Whether this is a problem depends on what the VM does next. If it has enough time to replenish its entropy pool, the state of the VMs will gradually diverge. But if it decides to generate a key: not good at all.

Embedded devices present their own class of problems. Unfortunately (like every other problem in the embedded arena) there’s no general solution. Some people obtain entropy from user keypad timings — if there is a user and a keypad. Some use the low-order bits of the ADC output. Still others forgo this entirely and ship their devices with an externally-generated PRG seed, usually stored in NVRAM.

I don’t claim that any of these are good answers, but they’re better than the alternative — which is to pretend that you have entropy when you don’t.

How do pseudo-random number generators work?

You’ve read the books. You’ve seen the movies. But when it comes down to it you still don’t understand the inner workings of the typical pseudo-random number generator. I can’t possibly make up for this in a single blog post, but hopefully I can hit a few of the high points.

Block cipher-based PRGs. One common approach to PRG construction uses a block cipher to generate unpredictable bits. This seems like a reasonable choice, since modern block ciphers are judged for their quality as pseudo-random permutations, and because most crypto libraries already have one lying around somewhere.

ANSI X9.31 PRNG implemented with AES (source). At each iteration, the PRNG takes in a predictable ‘date-time vector’ (DTi) and updated state value (Si). It outputs a block of random bits Ri. The generator is seeded with a cipher key (k) and an initial state S0.

One inexplicably popular design comes from ANSI X9.31. This PRG is blessed by both ANSI and FIPS, and gets used in a lot of commercial products (OpenSSL also uses it in FIPS mode). It takes in two seeds, k and S0 and does pretty much what you’d expect, on two conditions: you seed both values, and you never, ever reveal k.

If k does leak out, things can get ugly. With knowledge of k your attacker can calculate every previous and future PRG output from one single block of output!** This is totally gratuitous, and makes you wonder why this particular design was ever chosen — much less promoted.

Before you dismiss this as a theoretical concern: people routinely make stupid mistakes with X9.31. For example, an early draft of the AACS standard proposed to share one k across many different devices! Moreover keys do get stolen, and when this happens to your RNG you risk compromising every previous transaction on the system — even supposedly ‘forward-secure’ ones like ephemeral ECDH key exchanges. You can mitigate this by reseeding k periodically.

Hash-based PRGs. Many PRGs do something similar, but using hash functions instead of ciphers. There are some good arguments for this: hash functions are very fast, plus they’re hard to invert — which can help to prevent rewinding attacks on PRG state. Since there are zillions of hash-based PRGs I’ll restrict this discussion to a few of the most common ones:

  1. FIPS 186-2 (Appendix 3) defines a SHA-based generator that seems to be all the rage, despite the fact that it was nominally defined only for DSA signing. Windows uses this as its default PRG.
  2. Linux uses a hash-based PRG based on two variants of SHA.
  3. The non-FIPS OpenSSL PRG also uses a hash-based design. Like everything else in OpenSSL, it’s clearly documented and follows standard, well-articulated design principles.
Left: the Linux PRG (circa 2006). Right: the non-FIPS OpenSSL PRG.

Number-theoretic PRGs. The problem with basing a PRG on, say, a hash function is it makes you dependent on the security of that primitive. If a the hash turns out to be vulnerable, then your PRG could be as well.*** (Admittedly, if this happens to a standard hash function, the security of your PRG may be the least of your concerns.)

One alternative is to use a PRG that relies on well-studied mathematical assumptions for its security. Usually, you pay a heavy cost for this hypothetical benefit — these generators can be 2-3 orders of magnitude slower than their hash-based cousins. Still, if you’re down for this you have various choices. An oldie (but goodie) is Blum-Blum-Shub, which is provably secure under the factoring assumption.

If you like standards, NIST also has a proposal called Dual-EC-DRBG. Dual-EC is particularly fascinating, for the following three reasons. First, it’s built into Windows, which probably makes it the most widely deployed number-theoretic PRG in existence. Second, it’s slightly biased, due to a ‘mistake’ in the way that NIST converted EC points into bits.**** Also, it might contain a backdoor.

This last was pointed out by Shumow and Ferguson at the Crypto 2007 rump session. They noticed that the standard parameters given with Dual-EC could easily hide a trapdoor. Anyone who knew this value would be able to calculate all future outputs of the PRG after seeing only a 32-byte chunk of its output! Although there’s probably no conspiracy here, NSA’s complicity in designing the thing doesn’t make anyone feel better about it.

Shrinking generator.

The rest. There are many dedicated PRG constructions that don’t fit into the categories above. These include stream ciphers like RC4, not to mention a host of crazy LFSR-based things. All I can say is: if you’re going to use something nonstandard, please make sure you have a good reason.

How much entropy do I need?

The general recommendation is that you need to seed your PRG with at least as much entropy as the security level of your algorithms. If you’re generating 1024-bit RSA keys, the naive theory tells you that you need at least 80 bits of entropy, since this is the level of security provided by RSA at that key size.

In practice you need more, possibly as much as twice the security level, depending on your PRG. The problem is that many PRNGs have an upper bound on the seed size, which means they can’t practically achieve levels higher than, say, 256 bits. This is important to recognize, but it’s probably not of any immediate practical consequence.

I don’t care about any of this, just tell me how to get good random numbers on my Linux/Windows/BSD system!

The good news for you is that modern operating systems and (non-embedded) hardware provide most of what you need, meaning that you’re free to remain blissfully ignorant.

On most Unix systems you can get decent random numbers by reading from /dev/random and /dev/urandom devices. The former draws entropy from a variety of system sources and hashes it together, while the latter is essentially a PRG that seeds itself from the system’s entropy pool. Windows can provide you with essentially the same thing via the CryptoAPI (CAPI)’s CryptGenRandom call.

Care must be taken in each of these cases, particularly as your application is now dependent on something you don’t control. Many cryptographic libraries (e.g., OpenSSL) will run their own internal PRG, which they seed from sources like the above.

I’ve designed my own PRG. Is this a good idea?

Maybe. But to be completely honest, it probably isn’t.

If I seed my PRG properly, is it safe to use RSA again?

Yes. Despite the title of the recent Lenstra et al. paper, there’s nothing wrong with RSA. What seems to have happened is that some embedded systems didn’t properly seed their (P)RNGs before generating keys.

I’m sure there’s more to it than that, but at a high level: if you make sure to properly seed your PRG, the probability that you’ll repeat a prime is negligibly small. In other words, don’t sweat it.


* The security definition for a PRG is simple: no (computationally limited) adversary should be able to distinguish the output of a PRG from a sequence of ‘true’ random numbers, except with a negligible probability. An equivalent definition is the ‘next bit test’, which holds that no adversary can predict the next bit output by a PRG with probability substantially different from 1/2.

** Decrypting Ri gives you (Si XOR Ii), and decrypting DTi gives you Ii. You can now calculate Si by XORing the results. If you know DT{i-1} you can now compute R{i-1} and start the process over again. This was first noted by Kelsey, Schneier, Wagner and Hall in the context of an early version (X9.17). It works even if you only have a rough guess for the timestamp values — a pretty reasonable assumption, since some implementations specify a counter for the DT values.

*** It’s also important to be clear what security properties you’re relying on with a hash-based PRG. Most of the high-profile attacks on hash functions (e.g., MD5) focus on finding collisions; they’re not attacks on the pseudo-random nature of the outputs. In practice, this means you usually get lots of warning before a hash function becomes unsuitable for use in a PRG. Or maybe you won’t! Fun stuff.

**** Dual-EC is another fun example of NIST developing provably-secure looking protocols, but not actually including a security proof. This is particularly bizarre, because the only conceivable reason to use something as slow as Dual-EC is to gain this level of provable security. The generator is divided into two parts: the first generates pseudo-random EC points (this part is provable under the DDH assumption). The other part turns these points into bits. It’s the latter part that has the biasing flaw. Amusingly, the potential ‘backdoor’ wouldn’t be possible if the designers had built this part differently.

RSA keys: no insight whatsoever

I have a deadline coming up so (substantial) posting will be light this week.

For those of you who don’t read the New York Times, the big story of the week is this paper by Lenstra, Hughes, Augier, Bos, Kleinjung and Wachterlet:

Ron was wrong, Whit is right

We performed a sanity check of public keys collected on the web. Our main goal was to test the validity of the assumption that different random choices are made each time keys are generated. We found that the vast majority of public keys work as intended. A more disconcerting finding is that two out of every one thousand RSA moduli that we collected offer no security. Our conclusion is that the validity of the assumption is questionable and that generating keys in the real world for “multiple-secrets” cryptosystems such as RSA is significantly riskier than for “single-secret” ones such as ElGamal or (EC)DSA which are based on Diffie-Hellman.

Lots of people have written insightfully on this topic. See Dan Kaminsky’s post here, for example, or Thomas Ptacek’s excellent multi-part Twitter musing. (Update: much better, see Nadia Heninger’s explanation at the end of this post.)

There must be something wrong with me, because I find it almost impossible to draw any deep insight at all from this work. Don’t get me wrong: the paper itself is a fantastic piece of research; it sets a new standard for data analysis on public keys and certs. I hope we see more like it.

But what’s the takeaway? That two-key systems are insecure? That intelligence agencies have known this for years? Maybe. Whatever. The takeaway to me is that one (or more) RSA keygen implementations had a crappy RNG, or didn’t properly seed its PRG.

That’s really good to know about, but it isn’t the big news that the paper’s title would imply. It doesn’t have any implications for the use of RSA or any other cryptosystem. I’d sure like to solve the mystery of which implementations we need to look out for, and how to make sure this doesn’t happen again, but that’s literally the only thing I take away from this — so far.

I don’t mean to sound like a curmudgeon. Really, I want to believe. Please help me!

Update: Mystery solved! Nadia Heninger has a post at Freedom to Tinker explaining that most of these keys were generated by embedded devices, and that — through a parallel research effort — they actually know which devices. Once again extremely nice work. Even nicer than Lenstra et al., since it’s actually useful. (I can only imagine how Nadia and her team have been feeling the past two days, seeing ‘their’ result all over the New York Times. That’s responsible disclosure for you.)