Unbound version 1.0 was released with port randomization features. The same features that have been made available in the various patches by other vendors the CERT alert last month.
Now Dan Kaminsky disclosed more details during the the august 2008 Blackhat Conference in Las Vegas we release additional counter measures. These counter measures were previously withheld in order to minimize the risk of disclosing details about several variations of the attack through reverse engineering.
With the current set of counter measures added, Unbound offers state of the art protection against the attacks described by Kaminsky. However, state of the art counter measures will not provide full protection, not in Unbound nor in other software. Although DNSSEC is hardly deployed, it is currently the only mechanism known to deal with spoofing and other kinds of attacks on the DNS.
More details in the ways that Unbound protects against spoofing are below.
What is Cache Poisoning
Poisoning a DNS resolver refers to the act of inserting fake, often malicious data into the resolvers cache. This can cause website visitors to be redirected from the site (e.g. their banking site) they thought to visit to a different web site, for example a phishing site.
The basic approach of poisoning DNS queries is to send fake replies that pretend to come from the authority servers to the caching resolver. Every DNS query carries a random query Identifier (16 bit number). Only replies that contain the same number are accepted. In order for the resolver to accept the fake replies, the Identifier in the incoming packet needs to match that of the outstanding question. That is, the attacker has to guess a number of 16 bits in length.
One can calculate the time how long it takes to guess the 16 bits Identifier. a detailed calculation can be found online ( draft-ietf-dnsext-forgery-resilience). The document contains formulas with all the variables involved. The example given in the draft has a 4 Mb/s attack rate to get a 50% chance to inject the fake data.
It comes down to this: it takes a certain time, on average, to guess the right random sequence. The example in the reference argues that the Kaminsky exploit takes about 10 seconds to guess the 16 bit value in the identifier. This is confirmed by implementations of the exploit as well as various calculations on public mailinglists.
In order to extend the time by which a packet can be succesfully replaced we need more than the 16 bits of random number that the query Identifier provides us. That can be done by putting random numbers in other parts of the query, and checking if the server puts the same number in a reply, without changing the protocol. It is hard to add these extra random numbers without breaking interoperability, because the reply is only defined to contain a copy of the 16 bit ID value. Once an extra random number is copied into the reply, a fake reply must guess that number. Every extra bit that needs to be guessed, increases the time by a factor of 2. The goal is to add enough extra bits that the chance of poisoning becomes very low (on average).
Unbound implements a number of methods to add random bits. The most important means to add randomness is to vary the port numbers from which the question is asked, another means is to use a hack that randomizes unused bits in the query name. Unbound implements even more methods. In addition, Unbound is careful in what to accept as information that can be cached. These techniques are explained in more detail below.
Note however that the increase in the amount of bits does improve your chances to safely cross the road but a bad packet may still hit you.
Real protection, where you are not subject to the whims of chance, is achieved by using DNSSEC. DNSSEC uses digital signatures to protect the data. With DNSSEC there is no chance of poisoning, independent of the number of random bits used.
Unbound implements the DNSSEC standard as specified in the RFCs ( RFC4034, RFC4035 ). This means that it can act as a validator and can thus check the digital signatures attached in replies. Of course, the domain name owner must have inserted these digital signatures in the first place.
In the absence of DNSSEC, unbound attempts to provide very good security. Without digital signatures, randomisation and filtering are currently the only options. Below, a technical categorisation is made of the methods employed by unbound to protect unsigned data.
Unbound contains a component we call a 'scrubber'. This component takes care of certain checks, disallowing (removing) possibly malicious content.
By adding more random data, a spoofed reply has to guess more data to get through, lowering the chances of a successful poison attempt.
Additional security measures
These measures are mostly to prevent remote execution exploits.
So the default setup has a randomness of
16 bits port
2 bits destination address (estimated average).
For a total of 34 bits of randomness. Other implementations provide 16 bits (or less) unpatched, 26 bits for patches utilizing only 1024 ports and 32 bits for patches using the fully available port range (around 60k). Unbound has been utilizing the full port range of about 60.000 ports since the release of version 1.0.
With a careful setup, enabling capitalisation and source address randomisation Unbound provides
16 bits port
2 bits destination address (estimated average).
2 bits source address (estimated average).
8 bits capitalisation (estimated average).
in total 44 bits of randomness.
Sample config file items to enable this amount of randomness.
Time to infection
We take 10 seconds to infect an unpatched server with 50% chance as a baseline. The table below shows the time until a poison attempt is successful. The numbers are subject to being guesstimates. Better numbers may become available, either from the Blackhat presentation, or other sources. The bottom line is that adding randomness is a short term fix.
* : These are not enabled by default. The capitalisation has not been standardised, and could result in a small number of cases in slow or no answer. The source addresses need the operator to configure multiple addresses for the computer.
In the table above, the Bits column shows the number of random bits that are echoed in replies. The 50% chance column shows the length of time needed before an attack has a 50% chance of success (guessing the random numbers). The 5% chance column shows how long it takes before an attack has a 5% chance of inserting fake data.
Note: 60000 sockets not 65536 sockets used randomly for unbound is assumed in the table entries for unbound. Unbound avoids some port numbers for compatibility.
Also note that the table above assumes a fairly low bandwidth usage. If a large network capacity is available, say a botnet, and it can use 1000x more resources, then perhaps also the attack can be conducted 1000x faster.
In the meeting of the IETF dnsext working group successful poisoning attacks against an unpatched server in as little as 1/10 of a second were demonstrated easily ( demo results ), showing that much smarter things can be done than the dumb attack assumed for the numbers here. Calculations by members of the working group showed a near perfect chance for 6-8 seconds. This could move the figures to be less optimistic.
Keep in mind that the thousands of days shown for unbound with capitalisation and source addresses configured should not be taken as strong security. It is likely that some measures can be outsmarted. Or that these numbers are overly optimistic (see text above). And the 44 bits is an average. If an attacker can work out how to attack domains or queries with less protection, the the benefits may be partially lost. Thus, the large time listed for 44 bits should be taken as an indication that it is pretty good, but not invulnerable.
As stated earlier, the real solution is to use DNSSEC. DNSSEC makes this time table a non problem, because in all these cases DNSSEC can detect the forgery. Especially users in Brazil, Bulgaria, Puerto Rico and Sweden or people using these zones regularly, should consider turning on DNSSEC because the TLD zone is DNSSEC secured. Do consider using the DNSSEC capabilities in unbound.