Three computer bugs this year exposed passwords, e-mails, financial data, and other kinds of sensitive information connected to potentially billions of people. The flaws cropped up in different places—the software running on Web servers, iPhones, the Windows operating system—but they all had the same root cause: careless mistakes by programmers.
Each of these bugs—the “Heartbleed” bug in a program called OpenSSL, the “goto fail” bug in Apple’s operating systems, and a so-called “zero-day exploit” discovered in Microsoft’s Internet Explorer—was created years ago by programmers writing in C, a language known for its power, its expressiveness, and the ease with which it leads programmers to make all manner of errors. Using C to write critical Internet software is like using a spring-loaded razor to open boxes—it’s really cool until you slice your fingers.
Alas, as dangerous as it is, we won’t eliminate C anytime soon—programs written in C and the related language C++ make up a large portion of the software that powers the Internet. New projects are being started in these languages all the time by programmers who think they need C’s speed and think they’re good enough to avoid C’s traps and pitfalls.
But even if we can’t get rid of that language, we can force those who use it to do a better job. We would borrow a concept used every day in the physical world.
Obvious in retrospect
Of the three flaws, Heartbleed was by far the most significant. It is a bug in a program that implements a protocol called Secure Sockets Layer/Transport Layer Security (SSL/TLS), which is the fundamental encryption method used to protect the vast majority of the financial, medical, and personal information sent over the Internet. The original SSL protocol made Internet commerce possible back in the 1990s. OpenSSL is an open-source implementation of SSL/TLS that’s been around nearly as long. The program has steadily grown and been extended over the years.
Today’s cryptographic protocols are thought to be so strong that there is, in practice, no way to break them. But Heartbleed made SSL’s encryption irrelevant. Using Heartbleed, an attacker anywhere on the Internet could reach into the heart of a Web server’s memory and rip out a little piece of private data. The name doesn’t come from this metaphor but from the fact that Heartbleed is a flaw in the “heartbeat” protocol Web browsers can use to tell Web servers that they are still connected. Essentially, the attacker could ping Web servers in a way that not only confirmed the connection but also got them to spill some of their contents. It’s like being able to check into a hotel that occasionally forgets to empty its rooms’ trash cans between guests. Sometimes these contain highly valuable information.
Heartbleed resulted from a combination of factors, including a mistake made by a volunteer working on the OpenSSL program when he implemented the heartbeat protocol. Although any of the mistakes could have happened if OpenSSL had been written in a modern programming language like Java or C#, they were more likely to happen because OpenSSL was written in C.
Many developers design their own reliability tests and then run the tests themselves. Even in large companies, code that seems to work properly is frequently not tested for lurking flaws.
Apple’s flaw came about because some programmer inadvertently duplicated a line of code that, appropriately, read “goto fail.” The result was that under some conditions, iPhones and Macs would silently ignore errors that might occur when trying to ascertain the legitimacy of a website. With knowledge of this bug, an attacker could set up a wireless access point that might intercept Internet communications between iPhone users and their banks, silently steal usernames and passwords, and then reëncrypt the communications and send them on their merry way. This is called a “man-in-the-middle” attack, and it’s the very sort of thing that SSL/TLS was designed to prevent.
Remarkably, “goto fail” happened because of a feature in the C programming language that was known to be problematic before C was even invented! The “goto” statement makes a computer program jump from one place to another. Although such statements are common inside the computer’s machine code, computer scientists have tried for more than 40 years to avoid using “goto” statements in programs that they write in so-called “high-level language.” Java (designed in the early 1990s) doesn’t have a “goto” statement, but C (designed in the early 1970s) does. Although the Apple programmer responsible for the “goto fail” problem could have made a similar mistake without using the “goto” statement, it would have been much less probable.
We know less about the third bug because the underlying source code, part of Microsoft’s Internet Explorer, hasn’t been released. What we do know is that it was a “use after free” error: the program tells the operating system that it is finished using a piece of memory, and then it goes ahead and uses that memory again. According to the security firm FireEye, which tracked down the bug after hackers started using it against high-value targets, the flaw had been in Internet Explorer since August 2001 and affected more than half of those who got on the Web through traditional PCs. The bug was so significant that the Department of Homeland Security took the unusual step of telling people to temporarily stop running Internet Explorer. (Microsoft released a patch for the bug on May 1.)
There will always be problems in anything designed or built by humans, of course. That’s why we have policies in the physical world to minimize the chance for errors to occur and procedures designed to catch the mistakes that slip through.
Home builders must follow building codes, which regulate which construction materials can be used and govern certain aspects of the building’s layout—for example, hallways must reach a minimum width, and fire exits are required. Building inspectors visit the site throughout construction to review the work and make sure that it meets the codes. Inspectors will make contractors open up walls if they’ve installed them before getting the work inside inspected.
The world of software development is completely different. It’s common for developers to choose the language they write in and the tools they use. Many developers design their own reliability tests and then run the tests themselves! Big companies can afford separate quality–assurance teams, but many small firms go without. Even in large companies, code that seems to work properly is frequently not tested for lurking security flaws, because manual testing by other humans is incredibly expensive—sometimes more expensive than writing the original software, given that testing can reveal problems the developers then have to fix. Such flaws are sometimes called “technical debt,” since they are engineering costs borrowed against the future in the interest of shipping code now.
The solution is to establish software building codes and enforce those codes with an army of unpaid inspectors.
Crucially, those unpaid inspectors should not be people, or at least not only people. Some advocates of open-source software subscribe to the “many eyes” theory of software development: that if a piece of code is looked at by enough people, the security vulnerabilities will be found. Unfortunately, Heartbleed shows the fallacy in this argument: though OpenSSL is one of the most widely used open-source security programs, it took paid security engineers at Google and the Finnish IT security firm Codenomicon to find the bug—and they didn’t find it until two years after many eyes on the Internet first got access to the code.
Instead, this army of software building inspectors should be software development tools—the programs that developers use to create programs. These tools can needle, prod, and cajole programmers to do the right thing.
This has happened before. For example, back in 1988 the primary infection vector for the world’s first Internet worm was another program written in C. It used a function called “gets()” that was common at the time but is inherently insecure. After the worm was unleashed, the engineers who maintained the core libraries of the Unix operating system (which is now used by Linux and Mac OS) modified the gets() function to make it print the message “Warning: this program uses gets(), which is unsafe.” Soon afterward, developers everywhere removed gets() from their programs.
The same sort of approach can be used to prevent future bugs. Today many software development tools can analyze programs and warn of stylistic sloppiness (such as the use of a “goto” statement), memory bugs (such as the “use after free” flaw), or code that doesn’t follow established good-programming standards. Often, though, such warnings are disabled by default because many of them can be merely annoying: they require that code be rewritten and cleaned up with no corresponding improvement in security. Other bug–finding tools aren’t even included in standard development tool sets but must instead be separately downloaded, installed, and run. As a result, many developers don’t even know about them, let alone use them.
To make the Internet safer, the most stringent checking will need to be enabled by default. This will cause programmers to write better code from the beginning. And because program analysis tools work better with modern languages like C# and Java and less well with programs written in C, programmers should avoid starting new projects in C or C++—just as it is unwise to start construction projects using old-fashioned building materials and techniques.
Programmers are only human, and everybody makes mistakes. Software companies need to accept this fact and make bugs easier to prevent.
Simson L. Garfinkel is a contributing editor to MIT Technology Review and a professor of computer science at the Naval Postgraduate School.