Paper Reading Questions

For each paper, you should submit two text files via the Gradescope submission web site, as described below. The submission is due before lecture. (The Gradescope entry code for 6.5660 is 3JZ23X.)

Warning: You must have Javascript enabled to view a specific question.

Suppose slot_size is set to 16 bytes. Consider the following code snippet:

      char *p = malloc(256);
      char *q = p + 256;
      char ch = *q;

Explain whether or not baggy bounds checking will raise an exception at the dereference of q.

Suppose a program has a buffer overflow vulnerability which allows an attacker to overwrite a function pointer on the stack (which is invoked shortly after the buffer is overflowed). Explain whether or not an attacker is able to exploit the vulnerability if the same program is run under XFI.

What's the worst that could happen if one service in OKWS were to leak its 20-byte database proxy authentication token?

Would a Unix application running in the Unix environment described in the KeyKOS paper (i.e., KeyNIX) be susceptible to the confused deputy problem? Explain.

What are the principals that Java uses for access control?

List possible causes of false negatives (missed vulnerabilities) and false positives (reported problems that are not vulnerabilities) in the system described by the paper.

Is the descendant policy just as secure as the child policy for frame navigation? Either explain why it is so, or describe a concrete counter-example.

Two simple questions to make you think about this paper: 1) Why does the SSL 3.0 design accept SSL 2.0 connections? It is simpler to only accept SSL 3.0 connections, and it avoids the risk of rollback attacks described in Section 4.6. 2) The second figure of Section 4.3 shows the attack flow for deleting a change-cipher message. The fix requires an additional check. Where in the flow should SSL perform this check?

Table 1 lists key pinning through DNS as achieving better defense against detecting MITM attacks than key pinning based on client history. Give an example of an MITM attack that can DNS key pinning can detect but key-pinning based on history won't. (A description of DANE is here.)

Suppose that a web application developer wants to avoid the security pitfalls described in the ForceHTTPS paper. The developer uses HTTPS for the application's entire site, and marks all of the application's cookies as "Secure". If the developer makes no mistakes in doing so, are there still reasons to use ForceHTTPS? Explain why not, or provide examples of specific attacks that ForceHTTPS would prevent.

What is the worst that could happen if the private key of a user is stolen (i.e., becomes known to an adversary)? Similarly, what is the worst that could happen if the private key of a service is stolen? How should the compromised user or service recover? Think about possible vulnerabilities in the recovery process if the user or service key is known to an adversary.

Consider the implementation of a Spectre attack from Appendix A of the original Spectre paper. On the first call of readMemoryByte() from main(), if the attack registers a cache hit for mix_i on line 76, what is the likely value of mix_i?

What are some other situations where an adversary may be able to learn confidential information by timing certain operations? Propose some ideas for how an application developer might mitigate such vulnerabilities.

Could an adversary compromise a server running the system proposed in the paper without being detected?

Suppose an adversary steals a laptop that uses BitLocker disk encryption. In BitLocker's design, Windows has a key to decrypt the contents of the drive.

  1. What prevents the adversary from extracting this key from Windows?
  2. If the adversary cannot extract the key, what prevents him or her from simply using Windows to access files?

Sketch out the Resin filter and policy objects that would be needed to avoid cross-site scripting attacks through user profiles in zoobar. Assume that you have a PHP function to strip out JavaScript.

What are the technical risks and benefits of running an onion router Tor node (i.e., not just a client) on your machine?

Do you think a worm similar to Stuxnet could be designed to compromise Linux machines? What aspects of Linux or Windows design do you think make worms easier or harder to write?

What factors control the precision with which Vanish can make data unreadable after exactly time T?

In Table 1, what causes the secure deallocation lifetime to be noticeably larger (for some applications) than the ideal lifetime?

How could an adversary circumvent Backtracker, so that an administrator cannot pinpoint the initial intrusion point?

How does the proposed system deal with an adversary that tries to frame someone else for the denial-of-service attack by marking the attack packets they send in some way?

Given that CAPTCHAs can be solved quite cheaply, do you think that open web sites should continue using CAPTCHAs, switch to some other mechanism, or not use any mechanism at all (e.g., if you believe any mechanism will be cheap to break, like CAPTCHAs)? Explain your reasoning.

A browser cross-site scripting filter is a common client-side XSS prevention mechanism built into many modern browsers. Here's a brief description of what it does, in the words of Adam Barth, one of the creators of such filters, XSS Auditor: "Basically, the filter checks each script before it executes to see whether the script appears in the request that generated the page. If it finds a match, it blocks the script from executing. [...]". Do you think such a filter may be effective at detecting DOM-based (entirely client-side) cross-site scripting? Please explain.

The paper only mentions one potential false positives arising because of the use of regular expression. Explain why it is indeed a false positive.

Why is it necessary to treat innerHTML field assignments in a special way in the Gatekeeper analysis?

Security and performance are often at odds in computer systems. Do you feel that object views is a performant enough mechanism for everyday use?

What are some of the disadvantages of fast-propagating worms?

The paper discusses the possibility of using memory scanning to deal with the problems of obfuscation, encryption, and polymorphism. While memory scanning will enable signature-based detection, do you see any drawbacks of this approach?

JavaScript malware often uses a variety of environment detection techniques. One such technique is to check the version of the browser, plugins such as Adobe Acrobat or Flash, operating system, etc. before delivering an exploit deliberately designed for that platform and environment configuration, as illustrated by the pseudocode below.

if(browser-is-ie-6 && adober-flash-version==10.1){
This leads to more reliable, successful exploits for the attacker. Do you see how this pattern may lead to false negatives in a runtime detector?

The paper mentions that typical Android applications execute on top of a Java virtual machine. What is the role of Java in ensuring overall security?

Would it be reasonable to run TaintDroid to track what data applications may be exfiltrating from your phone at all times? Would it be reasonable to use TaintDroid to enforce policies like ``no application can send my IMEI to the Internet''? Explain why or why not, and what changes would be needed to make TaintDroid applicable, if not.

While privacy seems to be one clear benefit of client-side personalization, what are some of the disadvantages of it?

What are the disadvantages of using a human-readable, pseudonymous identifier for the user within a federated identity system, instead of a crypto key or a long string of hexadecimal numbers?

How could the operators of the spam value chain, studied in this paper, make it more difficult to repeat such studies in the future?

After reading this paper, propose some ideas for how you might improve the usability of securely accessing WebSIS (

Suppose you are building an online multi-person game. You are worried that a player can cheat in various ways by modifying the game software, since it runs on the player's own computer, or sending arbitrary network messages to your game server. What security properties could you get by using TrInc in your game (e.g., a trinket comes in the box when you buy a game)? What security problems cannot be solved with TrInc?

The authors of the Capsicum paper describe several strategies for how to use Capsicum in several applications (Section 4). How would you recommend using Capsicum in the different components of OKWS? Are there features missing from Capsicum that would have made it easier to build OKWS?

Suppose you are helping the developers of a complex web site at to evaluate their security. This web site uses an HTTP cookie to authenticate users. The site developers are worried an adversary might steal the cookie from one of the visitors to the site, and use that cookie to impersonate the victim visitor.

What should the developers look at in order to determine if a user's cookie can be stolen by an adversary? In other words, what kinds of adversaries might be able to steal the cookie of one of the visitors to, what goes "wrong" to allow the adversary to obtain the cookie, and how might the developers prevent it?

Note: an exhaustive answer might be quite long, so you can stop after about 5 substantially-different issues that the developers have to consider.

Why is it important to prevent access to scope objects?

Suppose an adversary discovers a bug in NaCl where the checker incorrectly determines the length of a particular x86 instruction. How could an adversary exploit this to escape the inner sandbox?

Two-factor Authentication (2FA) is commonly used to authenticate users. For example, MIT allows as a second factor a code sent via an SMS message to the user's cell phone. The FIDO standard describes a 2FA scheme using a USB dongle. Which scheme is more secure? Which scheme is more user friendly?

Based on the different schemes described in the paper, what do you think would be a reasonable choice for authenticating users in the following scenarios, and what trade-offs would you have to make:

  1. Logging in to a public Athena machine in a cluster.
  2. Checking your balance on a bank's web site via HTTPS from a private laptop.
  3. Accessing Facebook from an Internet cafe.
  4. Withdrawing cash from an ATM.

Which of the vulnerabilities described in this paper (A1 through A5) do you think could have been found with some kind of automated tool (such as fuzzing or program analysis) and what might such a tool look like?

Think about other applications that you run on your mobile phone. How might you apply Koi's techniques to help ensure privacy in these other applications? What other techniques could be useful?

Could large email providers, such as GMail, Yahoo Mail, or Hotmail, use ideas from SybilLimit to better detect spam email? What assumptions would they need to check?

First, ignoring range metadata, what constraint would KINT generate for the count variable in the code from Figure 3?

Second, how can you simplify the snippet of code in Figure 1 using the NaN integers as described in Section 7?

Steve Bellovin's ``A Look Back'' paper was published in 2004, over 10 years ago (and the paper itself is a retrospective on his earlier paper from 1989). Which of the security problems in the TCP/IP protocol suite described in Steve Bellovin's paper are still relevant today?

After you have read about Django's security mechanisms, think back to ``The Tangled Web''. What security pitfalls still remain for developers using Django? Could you extend Django to help developers avoid those pitfalls, in a style similar to Django's existing protections?

For the different parts of the browser state shown in Tables 1-3, what are the security implications of a "yes"? Consider both of the two threat models that the authors put forward for private browsing.

What do Dropbox developers gain from the obfuscation measures described in the paper? Could they have made it impossible for the authors to perform this kind of reverse-engineering?

Consider the following query:

  SELECT SUM(GREATEST(salary, 100)) FROM employees;

The GREATEST(a, b) function returns the larger of a and b, so the above query returns the sum of all salaries in the employees table, rounding up any salaries below 100 to 100.

How could CryptDB rewrite this query to execute over encrypted data, using the encryption schemes described in the paper?

For a BROP attack to succeed, the server must not rerandomize canaries after crashing. Suppose that, after a server crashes, it creates a new canary by SHA1-hashing the current gettimeofday() value. Is this new scheme secure?

KLEE uses a satisfiability (SAT/SMT) solver to implement symbolic execution. What would go wrong if KLEE did not use a SAT/SMT solver, and instead tried all branches? What would go wrong if KLEE just guessed randomly about what a symbolic value could be?

What kinds of security vulnerabilities are still possible in an Ur/Web application? One approach might be to keep the OWASP Top-10 list in mind as you are reading the Ur/Web paper, and consider whether Ur/Web's features can eliminate certain classes of bugs, or whether it's still possible to have vulnerabilities.

A note from the paper author: this paper is a draft of a camera-ready conference paper, and if you have any bug reports or suggestions about the paper, the author (Adam Chlipala, would appreciate your feedback!

Security engineering classes often focus on technological mechanisms such as cryptography or programming techniques (i.e., controls) to prevent security problems, but safety and biomedical engineering classes tend to focus on risk management to balance risks and benefits. Consider the situation of requiring fast emergency access to control an implanted medical device that must also remain secure. If the overarching goal is patient safety, how might your choice of security mechanisms differ from traditional computing contexts? How do we achieve both safety and security while balancing risks and benefits that ensure patient safety?

In Haven, can the untrusted operating system remap a virtual page of an enclave to a physical memory page that the operating system controls without the enclave noticing? (Explain briefly your answer.)

Can the untrusted operating system remap a virtual page of an enclave to a physical memory page that the operating system controls without the enclave noticing? (Explain your answer briefly.)

Each time EXE adds a branch constraint it queries STP to check that there exists at least one solution for the current path's constraints. What would go wrong if EXE did not use STP, and instead tried all branches? What would go wrong if EXE randomly selected a branch to check?

Read over the lab 4 assignment and tells us what we should cover that would help you most. That is, what question do you have about lab 4?

As you are reading this paper, think about what attacks Google is trying to prevent with each of the security measures described in this document. Submit a list of potential attacks that Google is worried about, approximately 1 per subsection in the document.

Suppose an adversary compromises the insecure operating system running on a user's desktop computer, and when the user tries to launch their secure document editor using Cloud Terminal, the adversary opens their look-alike document editor instead. How can a user determine that they are interacting with a ``fake'' application and not the real secure document editor? What prevents the adversary from fooling the user?

As you are reading the paper on secure messaging schemes, try to figure out: how does email stack up against the proposed criteria for messaging? How hard would it be to adopt the techniques used in the various messaging systems to improve email security? What makes email different, if anything?

Ben Bitdiddle has an iOS device that he uses to keep track of confidential notes. What are the different attack scenarios that Apple considers where an adversary tries to obtain Ben's notes, and what part of the iOS security architecture prevents those attacks?

This ZIP file contains two versions of a simple program. One is called rick, the other morty. They are generated from the same source code and should process the included sample input in the same way. However, one of the programs has a number of LAVA bugs injected. Can you figure out which? Can you construct inputs that demonstrate a difference in execution between rick and morty?

You might spend some time looking at the disassembly for each program, identifying bugs and trying to find inputs to trigger them. Or you might try fuzzing, e.g., using afl. We will go over the two versions and dissect the triggering inputs at lecture.

Describe what specific problem might arise if Bitcoin created new blocks much faster than its current design does---say, every 30 seconds.

In the SUNDR strawman described in section 3.1, each client signs the entire history of operations up to and including its operation. Suppose that the client were to sign just its operation, rather than the entire history. Could an attacker violate SUNDR's goal (fetch-modify consistency, as defined in section 3) in this alternative design?

In Komodo, what prevents the following two attacks:

  1. A malicious OS kernel tries to incorrectly build an enclave's page table, using the same physical page (PageNr) for two different virtual addresses in the enclave's memory (i.e., passing the same dataPg to two invocations of MapSecure with different va values).
  2. A malicious OS kernel tries to corrupt the enclave's code when first constructing the enclave: that is, when invoking MapSecure to load the enclave's executable code, it supplies a content page argument that contains a corrupted version of the enclave's binary.

WebAssembly's design is simple but effective at supporting a high-performance runtime: that is, making it possible to safely run WebAssembly code in a sandbox while still achieving high performance. To help you appreciate what makes WebAssembly good at this goal, consider three alternatives: Python code, C code, and machine code (e.g., x86 or ARM).

What makes it difficult to write high-performance sandboxed runtimes for each of these three alternatives?

In what situation (application, workload, context, etc) would it make sense to use each of the isolation approaches discussed in this paper (LXC, gVisor, Firecracker)?

What mistakes might a developer make in using RLbox that would lead to their application (e.g., the Firefox renderer) being vulnerable to attacks by a compromised sandbox (e.g., the libjpeg library)? Can these mistakes be mitigated mechanically (i.e., by extending RLbox, perhaps at the cost of additional overhead) or not, and why?

Suppose that an adversary gains control of a device driver isolated using KSplit. Is it possible for the adversary to gain control over the entire Linux kernel? In what situations could this happen, or why do you think this is impossible in KSplit?

Suppose that the participants of a Zoom meeting (which is using end-to-end encryption) see that some person left the meeting (no longer shows up in the participant list). Can that person still decrypt the content of the subsequent conversation in the meeting? Why or why not?

What is an example provided in the paper of an attack that was mischaracterized? That is, classified as a Spectre attack when it was actually Meltdown or vice versa? Briefly explain the attack, and why the authors argue that the attack was incorrectly classified.

Why do Kwon et al. argue that website fingerprinting attacks are more realistic in the case of deanonymizing users accessing hidden services, as opposed to deanonymizing users accessing arbitrary websites?

What is the basic mechanism through which the attacker in the paper deduces that a function at a particular address has been called, given only a trace of page numbers, but not complete addresses?

Suppose that the developer implements the PIN-protected backup HSM by comparing the supplied PIN code to the stored PIN code one digit at a time, from left to right, returning an error as soon as it finds a mismatched digit. Explain why this implementation should not satisfy the paper's IPR definition for the specification shown in Figure 2, and explain what part of the proof would fail.