Slashdot Mirror
Software Update Shuts Down Nuclear Power Plant
Garabito writes "Hatch Nuclear Power Plant near Baxley, Georgia was forced into a 48-hour emergency shutdown when a computer on the plant's business network was rebooted after an engineer installed a software update. The Washington Post reports, 'The computer in question was used to monitor chemical and diagnostic data from one of the facility's primary control systems, and the software update was designed to synchronize data on both systems. According to a report filed with the Nuclear Regulatory Commission, when the updated computer rebooted, it reset the data on the control system, causing safety systems to errantly interpret the lack of data as a drop in water reservoirs that cool the plant's radioactive nuclear fuel rods. As a result, automated safety systems at the plant triggered a shutdown.' Personally, I don't think letting devices on a critical control system accept data values from the business network is a good idea."
Must restart reactor to complete software installation.
[Yes] [No] [OMFG!]
I'd rather it shut itself down then suffer major failure.
Adds a whole new meaning to "Critical Update".
http://blindscribblings.com - Tasty pop-culture in conceptual fashion.
Personally, I am reassured that these reactors are designed to shut down at the drop of a hat. This is not a situation were fuck-ups should be masked, any discontinuity, however minor, really needs to be highlighted and dealt with immediately.
"I bless every day that I continue to live, for every day is pure profit."
"Vent radioactive gas? Venting gas prevents explosion. [Yes / No]"
Took this comment seriously, did you?
I wonder if they were using something like EPICS. I worked on a large experiment which used EPICS to control the system. Rebooting a machine would sometimes expose a problem with resources not being freed, eventually leading to a situation where data channels would read the 'INVALID/MISSING' value. The solution, as anyone who has worked on this sort of experiment will know, was to reboot more machines until the thing worked.
(I don't mean to complain about EPICS. It is very powerful and flexible... it's just that the version we used had these occasional hiccups.)
did it run Windows?
The only possible interpretation of any research whatever in the 'social sciences' is: some do, some don't
However useful a tip that may be, it has nothing to do with this incident. You clearly never even made it to the article summary, let alone the actual article.
"... when the updated computer rebooted, it reset the data on the control system, causing safety systems to errantly interpret the lack of data as a drop in water reservoirs that cool the plant's radioactive nuclear fuel rods. As a result, automated safety systems at the plant triggered a shutdown."
From that snippet alone, it stands to reason that _any_ reboot of the computer would have caused this reset in at the control system. Nor is this at all surprising; go reset any data collection system connected to controller software for any sort of industrial process and see if the controller doesn't receive spurious data.
To me this is an example of the automated system doing it's job. "Hark! I am a coolant reservoir monitor and I have reason to believe there may be a loss of coolant inventory. Time to trip the system."
The only safe way to update a system is a reboot, sure you CAN do some stuff on linux bsd etc to avoid having to reboot( hell this was probably running some unix derivative so it was probably possible to do the update without rebooting), but you wouldn't want to run the risk of introducing an unchecked bug by doing a live update. when your choices are:
a) high chance of accidentally shutting down a reactor harmlessly
b) small chance of fucking up a nuclear reactor
you'll always go for (a), if your sane.
IranAir Flight 655 never forget!
It says right in the EULA that it's not to be used in a nuclear power plant!
jX [ Make everything as simple as possible, but no simpler. - Einstein ]
While perhaps the system should be designed to behave differently, what happened here was a good thing. When things went wrong, rather than the reactor systems freaking out and doing random crap, they were properly designed to shift to a known safe state (i.e. Shut the hell down).
I write this type of software for a living so I know that having a computer on the business network connected to the control computers is a risk, bur that risk can be managed. The problem here is that the software update wiped out the nuclear control system data. This exposes two bad problems. First customers are always asking why they can't update their system while it is still running. We liken that to changing your tire while driving down the road. Secondly the software update did not respect the data in the nuclear control system and synchronized it to new initial data in the update on the other system! Not a good idea. In critical safety systems, you always practice an update before actually doing one.
Think about the cost associated with having and maintaining a completely hot-pluggable second control system. How much do you want your power bills to go up to pay for that? And what would be the point?
They have a perfectly adequate safety system that did exactly what it's supposed to do. It read confusing data and decided to shut the reactor down until a human came along and explained things satisfactorily. What's wrong with that? Aside from having the reactor offline for 48 hours, there was no other cost.
"... The move to SCADA systems boosts efficiency at utilities because it allows workers to operate equipment remotely."
Another proof that Homer Simpson was truly ahead of his time.
Are you mad, woman? You never know when an old calendar might come in handy. Sure, it's not 1985 now, but who knows what tomorrow will bring? -Homer
"Don't let fools fool you. They are the clever ones."
Don't forget about the now mutated sharks living in the coolant water growing frickin' laser beams on their heads.
As our way of thanking you for your positive contributions to Slashdot, you are eligible to disable Slashdot 2.0.
What exactly do you find frightening about an automatic safety system doing exactly what it's supposed to in response to unusual input?
Correct. It is not the better choice. In the foreseeable future, it is the only choice.
"GROSS NEGLIGENCE - Failure to use even the slightest amount of care in a way that shows Recklessness or willful disregard for the safety of others." - 'Lectric Law Library.
Yeah, those bastards, the way they used THE SLIGHTEST AMOUNT OF CARE in designing a system that shuts down in response to unexpected data so as to avoid RECKLESSNESS with the SAFETY OF OTHERS.
Secondly the software update did not respect the data in the nuclear control system and synchronized it to new initial data in the update on the other system! Not a good idea. In critical safety systems, you always practice an update before actually doing one.
I have no problem with a computer on the process control subnet reporting information to a computer on the business subnet.
I have a BIG problem with a computer on the business subnet being able to modify and corrupt data in a computer on the process control subnet.
"I can't dump data to the business side" is a reason to make a log entry and maybe sound a minor alarm. It's not a reason to shut down the reactor (unless the data is needed for regulatory compliance and the process control side isn't able to buffer it until the business side is working correctly.)
But if a business subnet computer can tamper with something as critical as a process control machine's idea of the level of coolant in a reservoir, it rings my "design flaw" alarms.
Is it ONLY able to reset it to "empty" as poorly-designed part of a communication restart sequence? Or could it also make the process control machine think the level was nominal when it WAS empty?
IMHO this should be examined more closely. It may have exposed a dangerous flaw in the software design.
Security flaws don't care if they're exercised by mischance or malice. If nothing else, this is a way to Dos a nuclear plant through a breakin on the business side of the net.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
This is why you keep the IT nerds away from the process network.
I've had a whole plant lose view of it's system because some well meaning retard in IT decided to push updates onto a SCADA system without qualifying the updates....... never had it KILL the control side of things though....well done whoever you were, you've done well.
Burma?
The problem here is that the system didn't shut down because it detected an error in the data collection system, instead it incorrectly detected a problem that did not in fact exist and then proceeded to take action. While the engineer in me is fairly certain that the system is designed to always fail to a safe state (as in, any automatic emergency response couldn't accidentally make things worse - at least not without raising all sorts of alarms), it is still concerning that internal control systems can be so effected by external servers.
In the article they mention that the system wasn't designed for security (since it was meant to be internal) - but this isn't a security issue at all! Any sort of system that relies upon other systems should be designed to assume failure can and will occur in other systems - that is not to say that it needs to verify/evaluate incoming data to make sure it is "good", but rather that it can tell the difference between receiving data (such as current water levels) and receiving no data at all (system failure). Once it has that it can ideally automatically switch to a backup system, or do what it did here and enter a fail-safe state (the difference being that it does so while pointing out the actual problem and not a incorrectly perceived problem in a different part of the system).
And a shutdown, while incovenient, is not a catastrophe. In fact, it speaks well for the plant's safety that it did automatically shut down when faced with bad data.
At least it did not turn it into a meltdown, so at least the safety features worked in the software.
That is definitely a glass half full, as opposed to empty.
He's trying to find an opportunity to bash Microsoft!
most freakouts surrounding nuclear power are based on 1960s technology. modern reactor designs, such as pebble bed reactors, are designed to be passively safe. that is, you can just walk away from them, doing nothing, and they will not release gas, go china syndrome, or anything else unsafe. older nuke tech requires active safety management: someone must always be on the job, making sure nothing f***s up. designing safety into nuclear reactor design from the philosophical ground up is the way of the future
intellectual property law is philosophically incoherent. it is your moral duty to ignore it or sabotage it
"In the game of life, someone always has to lose. To me, if life were fair, that someone would always be Oklahoma." -DKR
Agreed. That was good software design to assume a worst-case scenario when the sensors stopped sending in data. The alternative (sending pager alerts or something) would be far worse.
No, it has no reason to believe the coolant system has water. It's called FAIL SAFE; if I'm not quite sure, then fuck it, back off and shut the grid down and go MAKE SURE everything looks right.
The proper response of a nuclear cooling system to not knowing whether or not it's working correctly is not "let's keep running hot and see if more sample data comes across."
Support my political activism on Patreon.
Good enough evidence for me! Microsoft caused a nuclear meltdown! Quickly, to the Blogo-Sphere!
From the summary: If it's monitoring the primary control system then it seems to me like the machine would have to be on the control network. The real issue is why did the primary control system accept a reset from a monitoring system. It sounds like there's more than one bug to track down.
When our name is on the back of your car, we're behind you all the way!
I have quite a few Windows 2003 servers that haven't been rebooted since August 2006 when we upgraded our computer room to a small datacenter (we went from a single busline and a constantly breaking AC unit to dual UPS's powered by separate generators and dual chillers with separate condensers.) It's not like it's impossible to get good uptimes on Windows, the only servers we reboot on a regular basis are our Citrix servers due to some bad code on Citrix's part that leaks memory over time and our Oracle server due to a bug where 10gR2 pulls time from the deprecate ticks counter (the same one that used to crash Windows9x) which rolls over after ~42 days. Both of those are the result of poor third party coding, not bugs in Windows.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
I'd go just a bit further and say that it speaks well for the software coders. There are at least three ways to treat any 'out of bounds' condition. They chose to make sure that the safe action was chosen.
An area where that loosely controlled type of team work gets into trouble unless all coders treat data passed to their code, and from their code in the same uniform functional ways.
It also makes me wonder how the code will react to certain malicious software, should it get loose in the facility. If I were writing code to destroy a nuclear facility, it is how data is passed from one process to another that I would definitely attack as well as other vulnerable places.
It is sort of reassuring to have seen a failure result in a controlled shutdown rather than some other, more undesirable action.
Support NYCountryLawyer RIAA vs People
Scenario : System comes up. Things don't work quite right. Some configurations are tweaked and system is now working fine.
Reboot. The tweaked configurations happen to go away. No one remembers which ones they were. The system is b0rked for a while.
I would hope that isn't the case for that system, but I have seen it happen before.
Yeah, so when a sensor breaks and stops sending in data, it'll keep running like usual, with maybe a small error code in the background. Cause, you know, that's how we want nuclear fucking powerplants to work.
It's not a nuclear power plant, but still, my network...
I've set nagios up to monitor my network, and any los of signal is considered CRITICAL, not just a warning, but critical... and I need to know then.
I will not give in to the terrorists. I will not become fearful.
Safety control systems in the chemical industry have been used for 20+ years. These systems have: - redundant CPU modules (which can be hot plugged) - redundant IO modules (which can be hot plugged) - redundant communication systems - self diagnostics (can detect a failed output transistor) - internal diagnostics (CPU voting to detect failed CPU core) - standard algorithms for redundant transmitters Shutting down is the "safer option" however there is still risks (such as thermal stressing pipework). It is a lesser of the two evils problem. This stuff is bread & butter for the chemical industry, there are a number of control companies that refuse to deal with the nuclear industry due to the requirement for unlimited indementy. ZombieEngineer
The system as a whole *did* know the reading was bogus. The control/safety system shut down because it stopped getting "safe" indications from the monitoring/input system. It seems pretty clear that the input system itself correctly logged the reason for the error.
The interface to the control system for the tank level doesn't (or at least shouldn't) have an entire separate "error" parameter -- it probably takes a simple numeric value from the input system.
The input software knows when the reading are bogus or missing. In that case it either stops sending input, which would presumably trigger a watchdog in the control system, or it sends data that indicates a worst-case scenario. with which the control system can do whatever it does in a worst-case scenario.
The control system itself doesn't care why there is or may not be safe input parameters, it only cares that it cannot rely on the input it needs for safe operation. Giving it any more information just adds code and interface complexity to safety-critical software.
Here's the system as implemented:
level = tank.getLevel()
if (level < SANE_MIN || level > SANE_MAX)
level = 0
control.input.set(TANK_LEVEL, level)
Here's the system you describe:
error = 1
level = tank.getLevel()
if (level > SANE_MIN && level < SANE_MAX)
error = 0
control.input.set(TANK_LEVEL, level, error)
The later makes the safety-critical control software more complex, with more test cases and more input parameters, none of which add any value to the safe operation of the control system. The error parameter potentially allows for operation during transient errors, but that's a decision you can make in other ways, without adding interface complexity.
The only inconvenience of the simpler interface is that you have to check the logs from the input device in addition to the control device to determine why the error occurred. And please don't argue that consolidated error logging is worth extra code complexity -- that's probably not even true in a web app, let alone a human-safety control system.
You can't put too much water in a nuclear reactor.
You call this a sig?
That's the thing, though; it's a misuse of a phrase so much as "kick the bucket" as a literal expression of kicking a bucket would be a misuse of a phrase. The definition of it can be easily achieved by examining the words it uses and their contexts, making it much less likely to confuse a non-native speaker than many other expressions in wide use. The main source of confusion would be if someone tries to make it out to be an invalid phrase.
One of the great things about English is that one can phrase something a million different ways and still get the same meaning; banning the use of one phrase because it happens to also be the name of a logical fallacy is silly and pointless.
There are such requirements in the US, be they for SIL ratings, performing haz-op reviews, etc. Particularly in nuclear apps.
In a plant, not all control systems are SIL rated, but the safety backups usually are....though more and more operators are buying or upgrading to SIL qualified systems and extending SIL to other than just the safety and protection backups.
In this case, the engineers were probably asleep at the wheel and didn't realize the changes they made to the control software impacted the trip & protection systems, so didn't bother to even have a haz-op review prior to making the change to get updates to a control parameter (or set of parameters) from a networked device. They probably figured they were just adding a trim or tuning variable of some kind to the control loop and didn't do ANY real failure analysis.
Oops.
Oh well, time for all the governing bodies like the NRC to get out the microscopes and take a peek at the plant's operating procedures and engineers adherence to them.
Cheers
What part of FAIL SAFE don't you understand?
The System FAILED. It is programmed to SAFE the reactor when shit happens.
Without its sensors it had no choice but to assume worse case and scram the reactor.
It did it the right way. It did it the way it was programmed to do it.
What would you have it do to determine why it is no longer getting critical data? Send out a droid to check the cat5 cables? Its a frikin computer in a rack, not R2D2.
It worked the way it was supposed to.
Take a step back and let the big boys handle the reactor, Please.
Sig Battery depleted. Reverting to safe mode.
Well, the auditors seem to expect it... as do the vendors when we call for support - "Oh, you say foobar isn't working... well it looks like you're 15 revisions behind; why don't you just fix that and call me when you're done. Oh, your policies state you need to test and certify them? Well I guess I won't be hearing from you for a while, then."
--
.nosig
I think you're missing the real point, which is that the central safety systems are being fed data from a 'business network'. What would happen if that computer had an issue that caused it to send the same data continuously even when the coolant level had really dropped? WHY are any safety systems receiving data from an insecure network?
It's bad enough that most reactors use regular PC's to do the data collection and reporting, given the security risks posed by such systems (especially if networked), but I never realized they would be so stupid as to feed data in the other direction like this!
I think you're missing the real point, which is that the central safety systems are being fed data from a 'business network'. What would happen if that computer had an issue that caused it to send the same data continuously even when the coolant level had really dropped? WHY are any safety systems receiving data from an insecure network?
It's bad enough that most reactors use regular PC's to do the data collection and reporting, given the security risks posed by such systems (especially if networked), but I never realized they would be so stupid as to feed data in the other direction like this!
Obviously you have -zero- experience with power plant networks. Allow me to enlighten albeit anonymously.The reason machines like this receive data from networks that could be considered 'less secure' is because telemetry is required from a multitude of sources to actually ascertain any useful realtime information. Aggregation machines have to speak many different protocols and translate between them while communicating with other machines that belong at other plants, cities, states, and even companies to effectively get an accurate picture of the entire grid's current conditions.
The world of plant control machines themelves is very vendor-driven. Most facilities have turnkey solutions brought in by the few major players in this field. ABB, Hathaway, GE, etc. Those players don't even use the same SCADA protocols. Some use ICCP, some use DNP, and others prefer Etherpoll. I've seen RS232 data encapsulated into everything from fully-meshed TCP connections via OSI-Soft's PI to barely encoded into modbus and slapped onto ethernet with only an understanding of ARP.
The solutions are required because electricity is not just one powerplant pumping watts blindly. Instead, you have a multitude of plants all pushing power onto ISO-controlled grids that all have to work in concert with each other. This requires -- yes, you guessed it -- networking! The world of plant networks is pretty complex despite the hype you see in the media. The business of making actual watts appear magically at your house at a nice, consistent 60Hz is vastly more involved that most people realize.
Telemetry comes from secured networks, business networks, and other companies and controlling agencies. That is how it works. Period.
If you are actually interested in seeing the way these are regulated to be secured, the information is cleverly hidden in plain sight at the NERC website.
Simplicity is better than complexity if you're really after security. You could write a small Web server, which did nothing more than respond to HTTP requests, which was provable secure. It's been done. But it's also one more piece of software that has to be maintained. Or use a large Web server, such as Apache. It's been a long while since there was a remote exploit against Apache, when it was simply serving static pages. A DOS attack might still be possible, but that shouldn't accomplish anything but revealing the attack, as long as software running on the systems on the control LAN, which update the data host, don't become wedged if the data host becomes unresponsive. Which you would still want to test for, BTW.
*However*, one of the more powerful ideas in configuring highly secure LANS is that the more-secure LAN is simply never allowed to accept connections from the less-secure LAN. It's also something that's really easy to firewall, your network becomes easier to audit, etc. If you're a security practitioner, it makes your life easier. You still have to worry about the sneaker-net, physical security, etc., but now you're more able to focus your resources on those areas. Once again, simplicity is better than complexity if you're really after security.
I don't know where you got the idea that I thought it was, "sooo impossile to have a perl script or whatever fetch the webpage and cut out the data you care about." It's easy. But pretty much nothing is as easy to extract data from as a CSV file, which you could process with nothing more than awk. That doesn't get you far with automating report generation, populating a database, or whatever else you intend to *do* with the data, but there are endless tools for those jobs--Perl included.
Also, in my experience, people want to mess with Web pages. They're more visual, and people tend to want to 'improve' them, meaning your Perl screen-scraper likely has to change as well. I see a lot less clamor for changing the data format in CSV files.
In the end, use what you need--XML, for all I care. Just *don't allow your less-secure LAN to initiate connections into your more-secure LAN*. That was the root cause of the failure described in TFA. It's one of many reasons the rule is so basic, though obviously not yet widely-enough followed. Ideally, hosts on a secure LAN communicate with *nothing* outside that LAN. You justify and document every[1] step away from that ideal, if for no other reason than that it plays hell with formal trust models, which can be important inputs into designing a thorough audit. I don't see how you justify accepting incoming traffic when there's an easy way to avoid it. In an audit, I'd be busting you for that Web server. Simple as that.
An approach like the one above is likely to make life easier for several internal groups, including office staff. And quite possibly the ultimate users--power consumers.
[1] I mean every, not most. For example, how do you handle time? I favor an NTP server on the secure LAN taking time inputs from the GPS cloud. I've never worked for an organization that had a spare atomic clock lying around, or I'd have used that, and eliminated one more external data flow.
What you do with a computer does not constitute the whole of computing.