Just as there can be appsec truths, there can be appsec laws.
Science fiction author Arthur C. Clarke succinctly described the wondrous nature of technology in what has come to be known as Clarke’s Third Law (from a letter published in Science in January 1968):
Any sufficiently advanced technology is indistinguishable from magic.
The sentiment of that law can be found in an earlier short story by Leigh Brackett, “The Sorcerer of Rhiannon,” published in Astounding Science-Fiction Magazine in February 1942:
Witchcraft to the ignorant… Simple science to the learned.
With those formulations as our departure point, we can now turn towards crypto, browser technologies, and privacy.
The Latinate Lex Cryptobellum:
Any sufficiently advanced cryptographic escrow system is indistinguishable from ROT13.
Or in Leigh Brackett’s formulation:
Cryptographic escrow to the ignorant . . . Simple plaintext to the learned.
A few Laws of Browser Plugins:
Any sufficiently patched Flash is indistinguishable from a critical update.
Any sufficiently patched Java is indistinguishable from Flash.
A few Laws of Browsers:
Any insufficiently patched browser is indistinguishable from malware.
Any sufficiently patched browser remains distinguishable from a privacy-enhancing one.
For what are browsers but thralls to Laws of Ads:
Any sufficiently targeted ad is indistinguishable from chance.
Any sufficiently distinguishable person’s browser has tracking cookies.
Any insufficiently distinguishable person has privacy.
Writing against deadlines:
Any sufficiently delivered manuscript is indistinguishable from overdue.
Which leads us to the foundational Zeroth Law of Content:
Any sufficiently popular post is indistinguishable from truth.• • •
Here’s an HTML injection (aka cross-site scripting) example that’s due to a series of tragic assumptions that conspire to not only leave the site vulnerable, but waste lines of code doing so.
The first clue to the flaw lies in the querystring’s
stateparameter. The site renders the
statevalue into a
titleelement. Naturally, a first test payload for HTML injection would be attempting to terminate that element. If that works, then a more impactful followup would be to append arbitrary markup such as
<script>tags. A simple probe looks like this:
The site responds by stripping the payload’s
</title>tag and all subsequent characters. Only the text leading up to the injected tag is rendered within the
<HTML> <HEAD> <TITLE>abc</TITLE>
This seems to have effectively countered the attack. Of course, if you’ve been reading this blog for a while, you’ll suspect this initial countermeasure won’t hold up – that which seems secure shatters under scrutiny.
The developers worried that an attacker might try to inject a closing
</title>tag. Consequently, they created a filter to watch for such payloads and strip them. This could be implemented as a basic case-insensitive string comparison or a trivial regex.
And it could be bypassed by just a few characters.
Consider the following closing tags. Regardless of whether they seem surprising or silly, the extraneous characters are meaningless to HTML yet meaningful to our exploit because they belie the assumption that regexes make good parsers.
<%00/title> <""/title> </title""> </title id="">
After inspecting how the site responds to each of the above payloads, it’s apparent that the filter only expected a so-called “good”
</title>tag. Browsers don’t care about an attribute on the closing tag. They’ll ignore such characters as long as they don’t violate parsing rules.
Next, we combine the filter bypass with a payload. In this case, we’ll use an image
The attack works! We should have been less sloppy and added an opening
<TITLE>tag to match the newly orphaned closing one. A nice exploit won’t leave the page messier than it was before.
<HTML> <HEAD> <TITLE>abc</title id="a"> <img src=x onerror=alert(9)> Vulnerable & Exploited Information Resource Center</TITLE>
The tragedy of this flaw is that it shows how the site’s developers were aware of the concept of HTML injection exploits, but failed to grasp the underlying principles of the vuln. The effort spent blocking an attack (i.e. countering an injected closing tag) not only wasted lines of code on an incorrect fix, but instilled a false sense of security. The code became more complex and less secure.
The mistake also highlights the danger of assuming that well-formed markup is the only kind of markup. Browsers are capricious beasts. They must dance around typos, stomp upon (or skirt around) errors, and walk bravely amongst bizarrely nested tags. This syntactic havoc is why regexes are notoriously worse at dealing with HTML than proper parsers.
There’s an ancillary lesson here in terms of automated testing (or quality manual pen testing, for that matter). A scan of the site might easily miss the vuln if it uses a payload that the filter blocks, or doesn’t apply any attack variants. This is one way sites “become” vulnerable when code doesn’t change, but attacks do.
And it’s one way developers must change their attitudes from trying to outsmart attackers to focusing on basic security principles.• • •
You taught me language, and my profit on’t
Is, I know how to curse: the red plague rid you,
For learning me your language!
– Caliban (The Tempest, I.ii.363-365)
The announcement of the Heartbleed vulnerability revealed a flaw in OpenSSL that could be exploited by a simple mechanism against a large population of targets to extract random memory from the victim. At worst, that pilfered memory would contain sensitive information like HTTP requests (with cookies, credentials, etc.) or even parts of the server’s private key. Or malicious servers could extract similarly sensitive data from vulnerable clients.
In the spirit of Caliban, Shakespeare’s freckled whelp, I combined a desire to learn about Heartbleed’s underpinnings with my ongoing experimentation with the new language features of C++11. The result is a demo tool named Hemorrhage.
Hemorrhage shows two different approaches to sending modified TLS heartbeats. One relies on the Boost.ASIO library to set up a TCP connection, then handles the SSL/TLS layer manually. The other uses a more complete adoption of Boost.ASIO and its asynchronous capabilities. It was this async aspect where C++11 really shone. Lambdas made setting up callbacks a pleasure — especially in terms of readability compared to prior techniques that required binds and placeholders.
Readable code is hackable (in the creation sense) code. Being able to declare variables with
automade code easier to read, especially when dealing with iterators. Although hemorrhage only takes minimal advantage of the
unique_ptr, they are currently my favorite aspects following lambdas and
Hemorrhage itself is simple. Check out the README.md for more details about compiling it. (As long as you have Boost and OpenSSL it should be easy on Unix-based systems.)
The core of the tool is taking the
tls1_heartbeat()function from OpenSSL’s
ssl/t1_lib.cfile and changing the payload length — essentially a one-line modification. Yet another approach might be to use the original
tls1_heartbeat()function and modify the heartbeat data directly by manipulating the
s3->wrecdata via the
In any case, the tool’s purpose was to “learn by implementing something” as opposed to crafting more insidious exploits against Heartbleed. That’s why I didn’t bother with more handshake protocols or STARTTLS. It did give me a better understanding of OpenSSL’s internals. But still, I’ll add my voice to the chorus bemoaning its readability.• • •