Researchers Warn Computer Clocks Can Be Easily Scrambled Via NTP Flaws (networkworld.com)

alphadogg writes: Researchers at Boston University said this week that they've found flaws in the Network Time Protocol (NTP), a 30-year-old Internet protocol whose security shortcomings could undermine encrypted communications and even jam up bitcoin transactions. The importance of NTP was highlighted in a 2012 incident in which two servers run by the U.S. Navy rolled back their clocks 12 years, deciding it was the year 2000. Computers that checked in with the Navy's servers and adjusted their clocks accordingly had a variety of problems with their phones systems, routers and authentication systems. There is at least one alternative out there, and reason to use it.

Comments

Most NTP clients I've seen...

[Chris Mattern]

have an option for ignoring server updates if the time differential is too great.

Re:Most NTP clients I've seen...

[Penguinisto]

Not only that, but there are loads of other options, the most basic being that you can use your own personal stratum 1 device, and point all your servers to that. Hell, you can even build your own if you're truly paranoid.

IIRC by default the Windows AD infrastructure has member clients/servers using the Domain Controllers as their local source, and then you only need to re-point the DCs to whatever you want (the registry info documentation for NTP was a bit hinkey as late as Win7/Win2k8, so if you go mucking around in there, do it at your own risk). On the Linux/BSD/*nix side, there's a zillion options to beef up security, and they are drop-easy to enforce with any competent orchestration software (puppet, cfengine, chef, whatever). I only remember the Windows side because a previous employer had a distributed BI system that demanded that all component client devices (a mixture of Windows and Linux) must remain within 5 seconds of each other time-wise (else the whole thing threw an error and stopped).

Like sibling said though, most sysadmins don't dork around with NTP once they get a time source running - few will set up a local NTP relay of sorts, fewer still will have those sources use at least three different and vastly disparate sources to check against, and very few will set up a local authoritative stratum 1 box.

Lots of workarounds, and you really only have to set it up right once, with maybe an occasional (as in once-a-year-or-so) review and tuneup of the infrastructure.

Re:Most NTP clients I've seen...

[arglebargle_xiv]

There is at least one alternative out there

Whoever proposed tlsdate as an alternative to NTP has no idea how either NTP or tlsdate work. What moving to tlsdate is doing is replacing a well-designed clock-synchronisation protocol talking to precise time servers with an opportunistic gimme-whatver-time-you've-got mechanism that returns a one-off estimate of an approximate time on a web server, assuming the server doesn't just set the time field to random bytes as many do. They're totally different things.

If you're really worried about this, run your own stratum 1 clock and serve NTP off that. If you're worried about the cost of a dedicated NTP server, build it yourself using any number of instructions on the Internet, e.g. these ones.

Re:Most users I've seen...

[Anonymous Coward]

don't screw around with options once they get the thing initially functioning.

Re:Authenticated NTP?

[halivar]

Once an RFC is adopted by IETF (as the linked RFC is), it becomes a standard. Bro, do you even internet?

Ob

[Hognoxious]

systemd automatically knows what time it is, but it'll only tell you in binary.

Re:Eric Raymond rewrite

[0123456]

Isn't Eric Raymond in the middle of a major rewrite of the NTP software, with emphasis on security?

I heard it was going to become part of systemd.

So design things to not require synced clocks.

[tlambert]

So design things to not require synced clocks.

It's not like you couldn't include your idea of your local time (whatever it is) in your NFS requests, and then have the server take its idea of its local time, generate a delta, and apply that to all the timestamps that you are trying to set on a file. Or conversely, when you do a stat, the server could include its idea of the local time, the client could use that to generate a delta vs. its own idea of local time, and apply that delta to the time being reported up from the kernel to user space.

The whole idea of having to synchronize clocks between machines is rather moronic. When you have a billion mechanical computers wandering around in your body with robot bodies to e.g. fight a nasty cancer, do you really think there's going to be enough spare CPU cycles, RAM, or communications bandwidth for them to run NTP requests around to each other?

I recently fielded a request from someone who was building an embedded device; the trick was, it was going to be pre-programmed, then deployed everywhere, and not have local time beacons (i.e. it couldn't access local beacons, such as local cell towers, which send out "time is now" broadcasts). The question was: "How do I sync the time to the local time?".

My response was "Why?".

The reason finally boiled down to wanting to put the time in log files, and to display it on an LCD.

There was no reason for either of these: if the devices are Internet connected, just grab an HTTP header by hitting a known HTTP server, and log in UTC, since the time in the header will be reported as a UTC time + a zone delta. For the display: why the hell do you need to display the time on the small LCD? Because it was the only neutral thing he could think of to display on the LCD. "Can't I just look at my watch/iPhone/VCR/microwave/refrigerator/dishwasher/clock? Or just display it in UTC? Or display a PacMan animation instead of a clock?". "I guess so".

Problem solved with no need to sync clocks.

Non-synchronized clocks are only a problem if you let them be a problem/make them a problem.