The following article was pointed out to me by one of my students, Elpidoforos Economou (Source: Alitheia newspaper, Cyprus, 13 Nov 1995)
On Thursday 9 November a whole village (population 3000) had its telephones disconnected for nonpayment of their bills. According to the newspaper report, a "computer error" showing all subscribers in the village to be in default was to blame. Reconnection took place when the problem was discovered a few hours later.
[The following culled from a brief article in the Nov 4th Guardian (UK broadsheet)]:
The UK's two largest credit authorisation systems (Barclay's PDQ and NatWest's Streamline) failed on Saturday Oct 28th leaving retailers unable to verify customers cards.
In Barclay's case more than 40% of transactions failed due to "a bug in the system's software". For NatWest the problem was a huge backlog of calls [reason unstated in article] which delayed card authentication.
Both systems had contingency systems which allow retailers to telephone in authentication requests. However, due to the volume of sales, the lines rapidly became jammed.
A jaw-dropping item in the "Legal Grounds" column by Reynolds Holding in the 13 Nov 1995 *San Francisco Chronicle*:
A group of researchers at New York University Law School have been working on something called the "Solomon Project" to develop software to decide legal cases without bothering with juries.
Gee, where do I start? Are polygraph (so-called "lie-detector") results admissible in court? How about voice-stress analysis? Can it be fair to non-native English speakers? No appeals!?
A comment at the end of the column by trial consultant Karen Koonan sums it up: "It's pretty scary."George C. Kaplan firstname.lastname@example.org 1-510-643-5651
My earlier message reported an item in the "Legal Grounds" column in the S.F. Chronicle about the "Solomon Project" from New York University Law School. This is a grand and implausible scheme to decide legal cases by computer without bothering with a jury.
A couple of calls to NYU revealed:
A big-time software pirate was arrested in Los Angeles last week and charged with two felony counts of fraud and trademark violations. Authorities seized an estimated $1 million in illegally copied software, high-speed duplicating equipment and $15,000 in cash. Thomas Nick Alefantes, who calls himself "Captain Blood," allegedly sold his wares through advertising in trade publications and a mail order business. (Wall Street Journal 15 Nov 95 B10)
America Online issued a warning to its users about a destructive file attached to an e-mail message that has been circulating through its service and also over the Internet. The message itself is okay, but trying to run an attached "Trojan Horse" file called AOLGold or "install.exe" could crash a hard drive. (Atlanta Journal-Constitution 16 Nov 95 F7)
This note is intended as a break from the usual horror stories.
I just returned from Tokyo as a passenger on a non-stop flight from there to Newark. Twelve hours in the air in a Boeing 747-400.
Along the way I got a look at the flight plan and chatted with one of the flight crew about it. On this particular flight, we were filed with a destination of Detroit and with Cleveland as our alternate (All airline operations must state a primary and alternate airport; the alternate must be forecast to have weather substantially better than what is nominally required for landing. There must be enough fuel on board to fly to the primary airport, do a missed approach there, fly to the alternate, and land with 45 minutes of fuel still on board), with the intention of revising the flight plan enroute to have a destination of Newark and an alternate of Boston.
Why the revision? Because the fuel consumption estimates on which the flight plan is based have 10% padding. On a 12-hour flight, that means an extra 1.2 hours of fuel on board. That's a lot. So they plan to land in Detroit, which is an hour or so closer than Newark. Then, when they're over central Canada somewhere, they check again how much fuel they really have. If it's what they expect, the 10% padding is now a much smaller amount and, as expected, they have enough to continue to Newark. If, on the other hand, they can't prove that nothing is wrong, then they have a safe place to land with plenty of fuel on board.
I think that's a nice example of risk management and foresight.Andrew Koenig email@example.com
Glad to see mention of this -- but it's 2,000 *times* the current level of tapping rather than 2,000% thereof (Freeh's "correction" notwithstanding).
[The following message was inadvertently appended to RISKS-17.45, but you may not have seen it if your mail reader truncates digests after the normal digest terminator. Thus, it is replicated here, to make sure it gets noted and properly indexed. Sorry, Peter. Perhaps this was a serendipitous test to see if anyone reads RISKS to the end these days. Just a few of our long-time contributors did remark on it. PGN]There has been much recent discussion in RISKS of accidents and incidents involving the commercial fly-by-wire aircraft, the Airbus A319/320/321/330/340 series and the Boeing B777. I've collected the RISKS discussions since September 1993, along with synopses and a little commentary, in a hypertext compendium, entitled `Incidents and Accidents with Fly-by-Wire Commercial Airplanes', accessible through my home page
I intend that more RISKS and other discussion will appear as time goes by. Comments, additions and corrections welcome. If you have technical material that really should be linked, please let me know.Peter Ladkin
I forwarded the X-31 crash info to a friend (Martin Gomez firstname.lastname@example.org) who copied me the following correspondence. Steven Weller +1 415 390 9732 email@example.com
A friend forwarded your note on the X-31 crash, and it made me grind my teeth a little, so I thought I'd reply directly to you.
I am a private pilot, and I design flight control software for a living. Based on this experience, please allow me to disagree with a couple of points you made. 1) "It is apparent to me that this crash is the result of a bug in the flight control software and sensor hardware of the X-31 aircraft. The computer was unable to compensate for a loss of the airspeed indicator."
That is a non-sequitur. A "bug" is not the same as a missing feature. Microsoft Word 6.0 can't do my taxes for me, but that's not a bug. To me, a bug is when the computer does something other than what the designer intended. If the designer intended the computer to detect and compensate for the loss of a sensor, then you're right, it was a bug. My understanding, from talking to engineers on the X-31 project at Dryden, is that the flight computer was NOT meant to detect such failures. It therefore acted correctly, in the sense that it did what it was designed to do. Obviously, it would have been nice if it could detect that the airspeed sensor had failed, but it's inability to do that is not a "bug" as you suggest.
"(a reading of zero is pretty impossible while the aircraft is flying)"
That's certainly true, but I don't think that's what happened to the X-31. First of all, this aircraft had a Kiel tube, not a pitot tube. No, I don't know the difference either! Assuming they work under the same principle (the measurement of dynamic pressure) then, like pitot tubes, they are subject to the same two forms of icing: 1) the total pressure port gets clogged, or 2) the static pressure port gets clogged. You may already know all this, but ...
In case 1, the airspeed reading will not change as the airspeed changes. The air in the pitot tube (or Kiel tube, presumably) gets trapped. As the ALTITUDE changes, though, the static pressure will change, and the airspeed reading will decrease as the altitude decreases. In case 2, the indication will react correctly to changes in airspeed, but changes in altitude will have an odd effect: going lower will cause the airspeed to read higher, and vice-versa.
I heard that case 1 is what occurred to the X-31 that day. The pitot side iced up at altitude, and as the aircraft descended to return to base, the reading decreased, with the results you mentioned.
"We need to understand these differences before we use computers to take over the job of a human."
That's true, too. I assume your article (I'm unaware of the context, since it was forwarded) is a warning to software designers to be careful what they assume, etc. You must realize, though, that "it was a software problem" is the perennial cry of the hardware engineer who asked for the wrong software, and got what he asked for.
This is the reason that my company hired a pilot to design flight control software...I usually know what the computer "should" do. In most cases, though, software designers have to assume the specs they are given are valid, and take all failure modes into account. Is it a software problem that the X-31 can't land vertically? Or that it carries less than one hour of fuel? If a pilot crashes the remaining X-31 because he tried to land it vertically, or to fly coast-to-coast, that's not a software problem, so neither is failure to detect a sensor error.
[Guilty as charged. I agree with you, the X-31 crashed because of a SYSTEM DESIGN error, not a software bug. It would have taken more than a software change to prevent the loss of control.
I have some flight experience (1.5 hours solo before marriage and a house changed financial priority) and I have spent a couple of years writing flight navigation and autopilot software, as well as flight simulator software. I believe that my private pilot training had a significant impact on my ability to write autopilot code. I'm glad to hear that companies are recognizing that pilots write better flight control software (and this hopefully will extend to other disciplines, as well).
I didn't think through the complete implications of pitot tube icing (although I did remember that often the effect of clogging the dynamic pressure sensor is to cause airspeed indication to remain constant). It is an interesting point that IAS would gradually decrease as the altitude decreased. This would make for an interesting control problem. My copy of the Private Pilot Manual (Jeppesen Sanderson, 1990) notes a third possibility: If the pitot tube is blocked and its drain hole remains open, the airspeed reading will drop to zero (usually gradually). This failure mode may or may not be applicable to a Kiel tube.
The root problem remains: an instrument critical to the operation of the aircraft can fail in a way that is not detected by the system. Clogged airspeed indicators are a common problem, and they appear to cause catastrophic failure of the X-31. I see a recurring problem of software designers (and system designers) getting inadequate requirements specifications.
The real point is that we are "pushing the envelope" in using computers to control complex systems. When the systems are dynamic (like flight control) they are difficult to specify and difficult to test. We learn from our mistakes. Some mistakes have bigger consequences than others: our X-31 pilot was pretty lucky! Presumably, the next spec for the X-31 flight control system will have a line that reads: "failure of the airspeed indicator shall not cause loss of control of the aircraft", or "failure of ANY sensor shall not cause loss of control of the aircraft" (although I would NOT like to be on the team that had to test that latter requirement!).Andrew C. Fuller, E-Systems, ECI Div. Box 12248, St. Petersburg, FL 33733
Re: Faster computers will never make security safer! (Re: Palme, 17.45)
If you ignore issues other than the encryption (& assume it's RSA, say), and ignore what quantum computers might be able to do, then the difficulty of encrypting goes up with something like the cube of the size of the key, but the difficulty of cracking goes up with something like the exponential of the size of the key.
That's the whole reason encryption can work. If cracking were only a constant factor harder than encryption then the safety margin would never be enough. It has to actually scale differently--and (with the qualifications above) as far as we know it does.
The question of building encryption into all communications involves another scaling issue though: the speed and cost of communications vs. processors. For quite a while we've had 10Mbps ethernets while processors kept getting faster. If this were to continue, pretty soon we'd all be happy to let the processor spend a small part of its time encrypting all traffic. But if communication speed/cost catches up again then we'll still be tempted to send at least some things in the clear.--Steve
> The problem is that CPU time becomes less costly at the same ratio
> for the encrypter as for the encryption breaker.
This assumption is wrong.
Quite clearly an encryption algorithm that is as cheap to break as to encrypt is worthless -- encryption systems are by definition designed to be harder to break than to encrypt.
However Jacob was comparing *ratios*, and a less intuitive though demonstrable fact is that while increasing CPU speeds helps both encryptor and breaker, the performance boost helps the encryptor *more* than the breaker. This is because the encrypt and
functions are of different complexity.
For example the RSA public key cipher used for key exchange, when small public exponents are used the encryption process is of O(n^2), and decryption O(n^3), but the time to break is a function running in super polynomial time involving factoring a product of two large primes.
So, some estimates of time to break (factor RSA modulus using GNFS) (extracted from a post Bruce Schneier's made on cypherpunks):
# of bits mips-years required to factor
And (real figures - PGP) encrypt and decrypt for the same key sizes (100Mhz R4600 based Indy workstation ($5000 workstation)):
# of bits encrypt decrypt
1024 0.06 s 1.2 s
2048 0.11 s 7.7 s
So say for arguments sake that a particular encryption system was using 1024 bit keys, and that CPU speed increased by 10 fold.
Now the encryptor could use 2048 bit keys in place of 1024 bit keys, with the same (or slightly better) real time response.
The breaker however is faced with a 10^9 increase in MIPs-years required to break, and only a 10 fold increase in CPU to speed the attack. The breaker is 10^8 times worse off, and the encryptor is marginally better off.
As CPU speeds increase, the code breaker loses.Adam
In RISKS-17.45, Derek Lee Beatty comments about Microsoft Press's _Writing_Solid_Code_, specifically its claim that. "assertions that check
for programming errors [...] can be removed from the shipped versions for better performance." His comment is that this idea is "neither new nor so entertaining to RISKS readers, but it's sound advice."
Ouch. Hopefully most RISKS readers reacted immediately to this statement.
[See the next two messages. PGN]I'm fond of the saying that "When your program's internal state becomes invalid, the *best* thing that can happen is that it crashes immediately." Programmers make mistakes; failing an internal assertion will prevent a mistake from propagating through a program execution and leading to undetected severe execution errors. For example:
There are probably a few assertions which belong only in development code, because they are so expensive to check that they make the product unusable. In my experience, this is surely *not* the common case. If this is "the Microsoft way," please choose an alternate route.Bart Massey firstname.lastname@example.org
Derek Lee Beatty's support of Microsoft's code development is amazingly naive. Remove assertions that check for programming errors? Rigorous testing can not prove a program bug free, and removing these tests just makes the bug harder to find. The bug, when it finally manifests, may end up being blamed on a completely separate product...
This is of course to Microsoft's advantage. It lets them make claims like "there are no bugs in Microsoft products that the majority of users want fixed"... like the way Microsoft Project for Windows crashes under NT 3.5, or the way the POSIX subsystem in NT hangs unexpectedly...
(no, I haven't reported these to Microsoft. Why should I? They're not going to fix it... we're going to have to upgrade to 3.51 just to get around this bug, for all that Bill Gates claimed "A bug fix is the worst possible reason to release an upgrade" or words to that effect).
Suns maze of patches might be tough to figure out, but at least they admit they've got bugs and do their best to fix them. Bill Gates, ruler of the free world, remains in his state of denial.
(on another tack, the Java interpreter runs with no security enabled, for performance reasons. All the security is implemented in theorem provers and other test mechanisms prior to execution. This is much like the old Burroughs boxes, where the compiler guaranteed that no operation that violated the security policy was generated. It's much easier to do that than check the output of an untrusted compiler. I would much rather they implemented strict runtime checks at the interpreter level at the cost of some performance.)Peter da Silva, Bailey Network Management, 1601 Industrial Boulevard, Sugar Land, TX 77487-5013 +1 713 274 5180 (NIC: PJD2)
So now the book and Dekek Lee Beatty go hand in hand giving bad advice.
Solid code keeps the checks for programming errors! You can never ever guarantee that no programming errors are going to surface in a production version. Keeping the code that checks for "programming errors" is going to catch lots of errors earlier on.
Moreover it is going to catch errors that aren't programming errors. Continuing with an inconsistent internal data representation is a bad idea. Whatever the cause.
Off course you might remove some of the 'code to find bug xyz'. The internal consistency checks (This pointer shouldn't be "NULL", this "switch" needs no "default") should remain.Roger
Please report problems with the web pages to the maintainer