[cryptography] Weak keys in NetBSD
noloader at gmail.com
Mon Apr 1 06:40:31 EDT 2013
Due to a programming error, pseudorandom numbers supplied with a warning
of "insufficient entropy at creation" may only contain sizeof(int) bytes
(32 or 64 bits) of cryptographic randomness.
The first attempt to fix this bug, on January 26, 2013, contained a
different programming error with the same effect.
The NetBSD kernel uses an entropy pool to gather entropy from
system events, hardware random number generators, environmental
sensors, and other sources. This pool is never directly accessed
by user processes nor by most kernel subsystems. Instead, the
pool is used to key individual instances of a random stream
generator based on the NIST SP 800-90 CTR DRBG, per-consumer.
Each instance of the random stream generator itself mixes in local
data believed to contain additional entropy, or at least to be
difficult for an adversary to predict. As each stream generator
runs, some additional such data are mixed in each time data are output.
This additional entropy consists of kernel-private statistics such as
per-thread context switch count and ticks on-cpu, as well as cpu
cycle counter timestamps.
When the kernel boots, it creates several instances of the random
stream generator very early. Additional random number
generator instances may be created relatively early by, for example,
the sshd script generating keys when it first runs.
Because generators created very early may be keyed when the system
has very little entropy available, the system rekeys those generators
as soon as a "minimum entropy" threshold is passed, where the entropy
collection pool has enough bits to provide bits which are random in the
cryptographic sense: computationally infeasible to distinguish from
actual random noise.
However, internally, the pool has an "entropy estimator" which counts
how many bits have ever been put into the pool, for entropy consumers
who care about information-theoretic randomness rather than cryptographic
randomness. If a consumer asks for "GOOD" bits rather than "ANY" bits,
the consumer will get only entropy-estimate worth of bits -- no more.
The code for keying the stream generators which actually supply bits to
kernel and user consumers requests GOOD bits and then, if it does not
get as many as it wanted, and the user has indicated that cryptographic
randomness is sufficient (for example, by opening /dev/urandom rather
than /dev/random) requests ANY bits for the remainder.
Other random-number-generator bugs have resulted in generators' output
having a very small number of possible states. This bug is different:
the stream generator's output has the expected range of possible states,
but under some conditions, enough of the stream generator's input may
be guessable that it might be practical to reproduce its output by brute
The Original Bug
Due to a misplaced parenthesis, if insufficient GOOD bits were
available to satisfy a request, the keying/rekeying code requested
either 32 or 64 ANY bits, rather than the balance of bits required
to key the stream generator.
The result of this bug is that even after the minimum entropy
threshold was reached, the generators for in-kernel and userspace
consumers could in the worst case be keyed, or re-keyed, with as
few as 32 bits (on 32 bit platforms) or 64 bits (on 64 bit
platforms) of data, plus a 32-bit cycle counter value, plus the name
of the generator (an LWP ID for userspace, a fixed string for kernel
consumers), two 32-bit cycle counter snapshots, a count of context
switches, and a count of LWP ticks on-CPU, plus stack noise.
The most conservative estimate is that cryptographic keys produced
in this way contain only 32 bits (on 32-bit platforms) or 64 bits
(on 64-bit platforms) of entropy. A more optimistic estimate would
assume that from an attacker's point of view, additional entropy
(unpredictable bits, which must be guessed by brute force) are
contained in the two cycle counter values and the LWP statistics.
An additional consideration is that most consumers (including the
OpenSSL and OpenSSH key-generation functions and utilities) access
/dev/urandom via OpenSSL's RAND_bytes() function. This function
itself mixes in two addtional timestamp values (though these are
in seconds, not cycles), and also due to its implementation imposes
a variable delay between the two in-kernel cycle-counter
measurements mentioned above.
Nonetheless, we urge users to proceed according to the conservative
estimate, but provide the optimistic estimate in the interest of
The Mistaken Fix (code from 2013-01-26 until 2013-03-29)
To fix the original bug, multiple instances of the code for
keying stream generators were replaced by a single new function,
cprng_entropy_try(). This function took a flag argument indicating
whether the consumer wanted "hard" (information-theoretic) or "soft"
(cryptographic) randomness. If "hard" randomness was requested, the
function returned only GOOD bits, leaving the remainder of the
caller-supplied buffer untouched. If "soft" randomness was requested,
the function filled the rest of the caller-supplied buffer with ANY
bits. The function always returned a count of only as many bytes of
GOOD data as it could obtain.
The code calling the new cprng_entropy_try() function, however,
reversed the sense of the flag argument. This was missed by the
original author and in code review.
This could cause keying attempts originating from /dev/random device
node opens to receive buffers with some GOOD and some ANY bits, but
since the count of GOOD bits was correctly returned, the keying
operations were retried before any data were output, and the result
was merely less-efficient operation.
However, keying attempts originating from /dev/urandom device node
opens could receive only partially-filled buffers (with only as
many GOOD bits as were available). The code for this case did not
force an immediate rekey of the generator, but rather scheduled it
for rekey as soon as bits were available in the entropy pool (but
in any event, after its next use, since use of the generator invokes
the rekeying code).
Rekeying would occur with full entropy, but in the worst case, the
next generator output might contain *only* entropy from the
generator name (process ID) two cycle counter values, and
kernel-private LWP statistics. If output was through OpenSSL's
RAND_bytes function, additional entropy would come from that
functions' implementation as discussed above.
We do not know how to precisely quantify this entropy and urge
that any keys generated on NetBSD-current or NetBSD 6 systems
between 2013-01-26 and 2013-03-29 be regenerated, unless those
systems are known to have not been in an "insufficent entropy"
condition when the keys were generated (see discussion below).
The correct fix always overwrites the entire caller-supplied buffer
with output from the entropy pool and should be more robust against
Systems which never experience an "insufficient entropy" condition
(for example, systems with hardware random number generators supported
by NetBSD) are not impacted by this bug.
All cryptographic keys generated on NetBSD 6 or NetBSD-current (prior to
2013-03-29) systems should be regenerated, unless it is certain that
the system in question cannot have suffered a low-entropy condition
when the keys were generated.
In particular, since ECDSA ssh host keys are new in NetBSD 6 and
are generated by /etc/rc.d/sshd at system boot if not yet present, it
is likely that for systems that have been updated to NetBSD 6.0 or a
netbsd-6 branch kernel before the fix date, ECDSA host keys have been
considerably weakened by lack of actual randomness, especially since
with little system uptime stack contents will be more predictable than
For systems newly set up with NetBSD 6, all SSH host keys are suspect.
Other persistent cryptographic secrets (for example, SSH or SSL keys of
any type) generated using /dev/urandom on NetBSD 6 systems which may have
had insufficient entropy at key generation time may be impacted and should
Signature algorithms based on the discrete logarithm problem can
disclose the private signing key if the nonce 'k' is predictable.
However, DSA and ECDSA keys used (but not generated) on systems with
this bug are believed to be safe. This is because the OpenSSL
implementation of DSA signing feeds the digest of the message signed to
the OpenSSL random-number generator as ancillary input, with the result
that the nonce depends on the output of /dev/urandom, additional input
used to seed the OpenSSL RAND_bytes() generator, and the message itself,
and all bits from all inputs are well diffused across all bits output.
Solutions and Workarounds
Don't generate cryptographic keys, and don't use random numbers for
critical applications, on NetBSD 6 or NetBSD-current systems prior
to 2013-01-27, unless the system has sufficient "GOOD" entropy. In
practice, this means reading from /dev/random rather than /dev/urandom
when using system entropy to generate cryptographic keys.
Install a kernel containing the fix.
The fastest way to do that, if you are running or can run a standard kernel
built as part of the NetBSD release process, is to obtain the corresponding
kernel from the daily NetBSD autobuild output and install it on your system.
You can obtain such kernels from http://nyftp.netbsd.org/pub/NetBSD-daily/
where they are sorted by NetBSD branch, date, and architecture. To fix
a system running NetBSD 6.0 or the stable NetBSD 6.0 branch, the
most appropriate kernel will be the "netbsd-6-0" kernel. The "netbsd-6"
kernel can also be used; this is the branch that will become NetBSD 6.1.
To fix a system running NetBSD-current, the "HEAD" kernel should be used.
In all cases, a kernel from an autobuild dated 20130329 or newer must be
used to fix the problem.
If you cannot use the autobuilt kernels, then for all affected NetBSD
versions, you need to obtain fixed kernel sources, rebuild and install
the new kernel, and reboot the system.
More information about the cryptography