08 Sep 2013
Initial thoughts on the "BULLRUN" revelations
After reading through some – certainly not all, and admittedly not thoroughly – of the documents and analysis of the NSA “BULLRUN” crypto-subversion program, as well as various panicky speculation on the usual discussion sites, I can’t resist the temptation to make a few predictions/guesses. At some point in the future I’ll revisit them and we’ll all get to see whether things are actually better or worse than I suspect they are.
I’m not claiming any special knowledge or expertise here; I’m just a dog on the Internet.
Hypothesis 1: NSA hasn’t made any fundamental breakthroughs in cryptanalysis, such as a method of rapidly factoring large numbers, which render public-key cryptography suddenly useless.
None of the leaks seem to suggest any heretofore-unknown abilities that undermine the mathematics that lie at the heart of PK crypto (trapdoor functions). E.g. a giant quantum computer that can simply brute-force arbitrarily large keys in short amounts of time. In fact, the leaks suggest that this capability almost certainly doesn’t exist, or else all the other messy stuff, like compromising various implementations, wouldn’t be necessary.
Hypothesis 2: There are a variety of strategies used by NSA/GHCQ for getting access to encrypted communications, rather that a single technique.
This is a pretty trivial observation. There’s no single “BULLRUN” vulnerability; instead there was an entire program aimed at compromising various products to make them easier to break, and the way this was done varied from case to case. I point this out only because I suspect that it may get glossed over in public discussions of the issue in the future, particularly if there are especially egregious vulnerabilities that were inserted (as seems likely).
Hypothesis 3: Certificate authorities are probably compromised (duh)
This is conjecture on my part, and not drawn directly from any primary source material. But the widely-accepted certificate authorities that form the heart of SSL/TLS PKI are simply too big a target for anyone wanting to monitor communications to ignore. If you have root certs and access to backbone switches with suitably fast equipment, there’s no technical reason why you can’t MITM TLS connections all day long.
However, MITM attacks are still active rather than passive, and probably unfeasible even for the NSA or its contemporaries on a universal basis. Since they’re detectable by a careful-enough user (e.g. someone who actually verifies a certificate fingerprint over a side channel), it’s likely the sort of capability that you keep in reserve for when it counts.
This really shouldn’t be surprising; if anyone seriously thought, pre-Snowden, that Verisign et al wouldn’t and hadn’t handed over the secret keys to their root certs to the NSA, I’d say they were pretty naive.
Hypothesis 4: Offline attacks are facilitated in large part by weak PRNGs
Some documents allude to a program of recording large amounts of encrypted Internet traffic for later decryption and analysis. This rules out conventional MITM attacks, and implies some other method of breaking commonly-used Internet cryptography.
At least one NSA-related weakness seems to have been the
pseudorandom number generator specified in NIST SP 800-90; it was a
bit hamhanded as these things go because it was discovered, but it’s
important because it shows an interest.
It is possible that certain “improvements” were made to hardware RNGs, such as those used in VPN hardware and also in many PCs, but the jury seems to be out right now. But compromising hardware makes somewhat more sense than software, since it’s much harder to audit and detect, and it’s also harder to update.
Engineered weaknesses inside [P]RNG hardware used in VPN appliances and other enterprise gear might be the core of NSA’s offline intercept capability, the crown jewel of the whole thing. However, it’s important to keep in mind Hypothesis 2, above.
Hypothesis 5: GCC and other compilers are probably not compromised
It’s possible, both in theory and to some degree in practice, to compromise software by building flaws into the compiler that’s used to create it. (The seminal paper on this topic is “Reflections on Trusting Trust” by Ken Thompson. It’s worth reading.)
Some only-slightly-paranoids have suggested that the NSA and its sister organizations may have attempted to subvert commonly-used compilers in order to weaken all cryptographic software produced with them. I think this is pretty unlikely to have actually been carried out; it just seems like the risk of discovery would be too high. Despite the complexity of something like GCC, there are lots of people looking at it from a variety of organizations, and it would be difficult to subvert all of them while harder still to insert an exploit that would have been completely undetected. In comparison, it would be relatively easy to convince a single company producing ASICs to modify a proprietary design. Just based on bang-for-buck, I think that’s where the effort is likely to have been.
Hypothesis 6: The situation is probably not hopeless, from a security perspective.
There is a refrain in some circles that the situation is now hopeless, and that PK-cryptography-based information security is irretrievably broken and can never be trusted ever again. I do not think that this is the case.
My guess – and this is really a guess; it’s the thing that I’m hoping will be true – is that there’s nothing fundamentally wrong with public key crypto, or even in many carefully-built implementations. It’s when you start optmizing for cost or speed that you open the door.
So: if you are very, very careful, you will still be able to build up a reasonably-secure infrastructure using currently available hardware and software. (‘Reasonably secure’ meaning resistant to untargeted mass surveillance, not necessarily to a targeted attack that might include physical bugging: that’s a much higher bar.) However, some code may need to be changed in order to eliminate any reliance on possibly-compromised components, such as hardware RNGs / accelerators that by their nature are difficult to audit.
Large companies that have significant investments in VPN or TLS-acceleration hardware are probably screwed. Even if the gear is demonstrably flawed, look for most companies to downplay the risk in order to avoid having to suddenly replace it.
Time will tell exactly what techniques are still safe and which aren’t, but my WAG (just for the record, so that there’s something to give a thumbs-up / thumbs-down on later) is that TLS in FIPS-compliance mode, on commodity PC hardware but with hardware RNGs disabled or not present at both ends of the connection, using throwaway certificates (e.g. no use of conventional PKI like certificate authorities) validated via a side-channel, will turn out to be fairly good. But a lot of work will have to be invested in validating everything to be sure.
Also, my overall guess is that neither the open-source world or the commercial, closed-source world will come out entirely unscathed, in terms of reputation for quality. However, the worst vulnerabilities are likely to have been inserted where there were the least number of eyes looking for them, which will probably be in hardware or tightly integrated firmware/software developed by single companies and distributed in “compiled” (literally compiled or in the form of an ASIC) form only.
As usual, simpler will turn out to be better, and generic hardware running widely-available software will be better than dedicated boxes filled with a lot of proprietary silicon or code.
So we’ll see how close I am to the mark. Interesting times.
This entry was converted from an older version of the site; if desired, it can be viewed in its original format.