Report on hacking blind

Report on hacking blind

• 1.
  1 INDEX Chapter Title PageNo. Abstract 7 1. Introduction 8 2. History of buffer overflows 9 3. ROP tutorial 10 4.Buffer overflows today 11 5. BROP environment 12 6. Attack outline 13 7. Stack reading : ASLR de-randomization 14 8.BROP attack : 8.1 The pieces of the puzzle 8.2 Finding gadgets and stop gadgets 8.3 Identifying gadgets 8.4 Finding the procedure linking table 8.5 Controlling rdx via strcmp 8.6 Finding write 8.7 Concluding attack 8.8 Attack summary 8.9 First principles of attack 15 9. Implementation 20 10. Evaluation : 10.1 Performance 10.2 Stability 10.3 Source code aid 22 11. BROP prevention 11.1. Re-randomization 11.2. Step on crash 11.3. ROP protections 25
• 2.
  2 11.4. Compiler techniques 12.Related work 27 13. Conclusion 28 References 29
• 3.
  3 List of Figures Figure No. FigureDescription Page No. 1 BROP in different OSes 27 2 BROP beyond stack overflows 27 4 Variance in binaries 28 5 Remote fuzz testing 28
• 4.
  4 ABSTRACT We show thatit is possible to write remote stack buffer overflow exploits without possessing a copy of the target binary or source code, against services that restart after a crash. This makes it possible to hack proprietary closed-binary services, or open-source servers manually compiled and installed from source where the binary remains unknown to the attacker. Traditional techniques are usually paired against a particular binary and distribution where the hacker knows the location of useful gadgets for Return Oriented Programming (ROP). Our Blind ROP (BROP) attack instead remotely finds enough ROP gadgets to perform a write system call and transfers the vulnerable binary over the network, after which an exploit can be completed using known techniques. This is accomplished by leaking a single bit of information based on whether a process crashed or not when given a particular input string. BROP requires a stack vulnerability and a service that restarts after a crash. The attack works against modern 64-bit Linux with address space layout randomization (ASLR), no-execute page protection (NX) and stack canaries.
• 5.
  5 Chapter 1: INTRODUCTION Attackers havebeen highly successful in building exploits with varying degrees of information on the target. Open-source software is most within reach since attackers can audit the code to find vulnerabilities. In an effort to understand an attacker’s limits, we pose the following question: is it possible for attackers to extend their reach and create exploits for proprietary services when neither the source nor binary code is available? To answer this question we start with the simplest possible vulnerability: stack buffer overflows One advantage attackers often have is that many servers restart their worker processes after a crash for robustness. Notable examples include Apache, nginx, Samba and OpenSSH. We present a new attack, Blind Return Oriented Programming (BROP), that takes advantage of these situations to build exploits for proprietary services for which both the binary and source are unknown. The attack is enabled by two new techniques: 1) Generalized stack reading: this generalizes a known technique, used to leak canaries, to also leak saved return addresses in order to defeat ASLR on 64-bit even when Position Independent Executables (PIE) are used. 2) Blind ROP: this technique remotely locates ROP gadgets. The BROP attack enables robust, general-purpose exploits for three new scenarios: 1) Hacking proprietary closed-binary services. One may notice a crash when using a remote service or discover one through remote fuzz testing. 2) Hacking a vulnerability in an open-source library thought to be used in a proprietary closed-binary service. A popular SSL library for example may have a stack vulnerability and one may speculate that it is being used by a proprietary service. 3) Hacking an open-source server for which the binary is unknown. This applies to manually compiled installations or source-based distributions such as Gentoo. Benefits are: 1) A technique to defeat ASLR on servers (generalized stack reading). 2) A technique to remotely find ROP gadgets (BROP) so that software can be attacked when the binary is unknown. 3) Braille: a tool that automatically constructs an exploit given input on how to trigger a stack overflow on a server. 4) The first (to our knowledge) public exploit for nginx’s recent vulnerability, that is generic, 64-bit, and defeats (full/PIE) ASLR, canaries and NX.
• 6.
  6 Chapter 2: HISTORY OFBUFFER OVERFLOWS Buffer overflows are a classic vulnerability with a long history of exploits . Conceptually, they are relatively easy to attack. For instance, a vulnerable program might read data from the network into a buffer. In the early days of stack buffer overflows, it was common for an attacker to include malicious code as part of the payload used to overflow the buffer. As a result, the attacker could simply set the return address to a known location on the stack and execute the instructions that were provided in the buffer. An innovative technique, known as return-oriented programming (ROP), was developed to defeat defences based on non-executable memory. It works by linking together short code snippets already present in the program’s address space. Such code snippets, called gadgets, can be combined to form arbitrary computation. As a result, attackers can use ROP to gain control of programs without any dependence on code injection. Simpler variations of ROP are sometimes possible. ASLR creates a major challenge for attackers because it makes the address locations of code (or even the stack) impossible to predict in advance. Unfortunately, on 32-bit platforms, ASLR is constrained by the number of available bits (usually 16) for randomization. As a result, brute-force attacks can be quite effective. However, on 64-bit platforms there are typically too many random bits for brute-forcing to be feasible. In such cases, ASLR can still be circumvented, but only when combined with a vulnerability that leaks information about the address space layout, such as a format string . Canaries are another common defence against buffer overflow attacks. Canaries cannot prevent buffer overflows, but they can detect them retroactively and terminate the program before an attacker can influence control flow. Fig 2.1: ROP version of socket reuse shellcode. Gadgets are chained by placing their addresses on the stack. Value is checked to make sure it has not changed. This can prevent stack buffer overflows from being exploited because an attacker must correctly overwrite the secret value in order for the program to actually use an overwritten return address.
• 7.
  7 Chapter 3: ROP TUTORIAL Modernexploits rely heavily on ROP. The goal of ROP is to build an instruction sequence that typically spawns a shell (shellcode) based on existing code fragments (gadgets). Once a shell is executed, the attacker can execute more commands to continue the attack. Traditionally, exploits would inject off-the-shelf shellcode into the process and execute it. Figure 1 shows typical shellcode that pipes the attacker’s socket to standard input, output and error and executes a shell. Of course injecting shellcode is no longer possible because these days writable memory (e.g., the stack) is non-executable, and so ROP must be used instead. Figure 2 shows how ROP can in principle be used to create the shellcode previously shown in Figure 1. The stack is overflowed so that the addresses of all the gadgets are present in sequence. Each gadget ends with a return so that the next gadget can execute. In practice, each ROP gadget will be a short sequence of machine instructions terminated by a return. Executing a simple system call like dup2 will require multiple gadgets , Fig 3.1 : ROP chain for dup2(s, 0). The systemcall number needs to be in rax and dup2 is system call #33. Arguments are passed in rdi and rsi. Because the attacker already controls the stack, pop gadgets are used to load registers from values on the stack. because arguments are passed in registers so gadgets to populate these will be needed. Figure 3 shows the required gadgets for dup2. Registers rdi and rsi control the first two arguments to system calls, and rax controls the system call number. Registers can be controlled by using pop gadgets and placing the value to load on the stack. By chaining enough gadgets, complete shellcode can eventually be built.
• 8.
  8 Chapter 4: BUFFER OVERFLOWSTODAY On most contemporary operating systems, where NX and ASLR are common, an attacker must fulfill at least two requirements in order to gain full control of a remote program’s execution: 1) To defeat NX, the attacker must know where gadgets reside inside the program executable. 2) To defeat ASLR, the attacker must derandomize the location at which the executable’s text segment is actually loaded in memory. These requirements can easily be brute-forced on 32- bit systems , through simple guessing. the rest of this discussion exclusively considers 64-bit attacks. The first requirement in practice means that the attacker must have a copy of the vulnerable binary to disassemble and find gadgets. To our knowledge, our proposed BROP attack is the first general-purpose technique that can be used to defeat NX when the binary code is completely unavailable. Defeating ASLR is also a significant challenge without BROP, but there are some possible strategies. Firstly, an information leak might reveal the address location of a code segment. Secondly, it may be possible to exploit any code that remains statically positioned across executions. As a result, an attacker could simply apply ROP to the program’s text segment directly. Fig4.1 : Techniques needed to attack different 64-bitscenarios.BROP and stack read are our contributions. With GCC, this is achieved through the -pie flag. Using PIE has been recommended in previous studies , but unfortunately, it has not been widely deployed to date. When PIE is enabled, there are no known general-purpose 64-bit techniques, outside of our proposed generalized stack reading attack, that can be used to defeat ASLR. Figure 4.1 shows how our BROP attack improves the state of the art in 64-bit exploit techniques.
• 9.
  9 Chapter 5: BROP ENVIRONMENT TheBlind Remote Oriented Programming (BROP) attack makes the following assumption and requires the following environment: • A stack vulnerability and knowledge of how to trigger it. • A server application that restarts after a crash. The threat model for a BROP attack is an attacker that knows an input string that crashes a server due to a stack overflow bug. The attacker must be able to overwrite a variable length of bytes including a return instruction pointer. The attacker need not know the source or binary of the server. The attacker is able to crash the server as many times as he wishes while conducting the attack, and the server must restart. Fig 5.1: Stack reading. A singlebyte on the stack is overwritten with guess X. If the service crashes,the wrong valuewas guessed. Otherwise, the stack is overwritten with the same valueand no crash occurs.After at most 256 attempts, the correct valuewill beguessed. The process is then repeated for subsequent bytes on the stack. If the server is compiled with the PIE flag, the server must be a forking daemon and must restart without using execve. The same is true for overflows where the canary must be modified by the exploit. The attacker is also able to distinguish when a server crashes prematurely, e.g., by noticing that the socket closes without receiving a response.
• 10.
  10 Chapter 6: ATTACK OUTLINE TheBROP attack has the following phases: 1) Stack reading: read the stack to leak canaries and a return address to defeat ASLR. 2) Blind ROP: find enough gadgets to invoke write and control its arguments. 3) Build the exploit: dump enough of the binary to find enough gadgets to build a shellcode, and launch the final exploit. The first phase is needed so that a starting point address for scanning gadgets is found. Gadgets are then searched for until enough are found to invoke write. After that, the binary is transferred over the network from memory, enabling known techniques to be applied toward building the final exploit.
• 11.
  11 Chapter 7: STACK READING:ASLR DE-RANDOMIZATION Exploits must have a method of defeating ASLR for configurations where PIE is used. We present a new stack reading technique that generalizes a known technique used for leaking canaries. It is useful even in cases where the binary is known and a full BROP attack is not required. The basic idea in leaking canaries is to overflow a single byte, overwriting a single byte of the canary with value x. If x was correct, the server does not crash. The algorithm is repeated for all possible 256 byte values until it is found (128 tries on average). The attack continues for the next byte until all 8 canary bytes (on 64-bit) are leaked. Figure 5.1 illustrates the attack. After the canary, one typically finds the saved frame pointer and then the saved return address, so three words need to be read. Figure 7.1 shows a typical stack layout. There are a few subtleties that apply to generalized stack reading but not to reading canaries. With canaries, exact values Fig 7.1 : showsa typical stacklayout. will always be returned because there is only one value canary that is correct. In general though, stack reading will not necessarily return the exact saved instruction pointer present on the stack. It is possible that a slightly different value is returned depending on whether another value still resumes program execution without causing a crash It is possible that stack reading does not return an address in the application’s own .text segment, but rather a return address in a library. On 64-bit x86 systems only a portion of the address space is made available to the operating system (canonical form addresses). This allows us to skip several bytes when reading pointers. For user space processes the top two bytes are always zero. In fact, on Linux the third byte is 0x7f for libraries and the stack. The main binary and heap are usually stored at 0x00 for executables compiled without the PIE flag. Thus we can skip on average three bytes (384 requests) when reading addresses. The words returned by stack reading give further evidence of the BROP attack working because the values can be somewhat sanitized.
• 12.
  12 Chapter 8: BROP ATTACK TheBROP attack allows writing exploits without possessing the target binary. It introduces techniques to find ROP gadgets remotely and optimizations to make the attack practical. 8.1. The pieces of the puzzle The goal is to find enough gadgets to invoke write. After that, the binary can be dumped from memory to the network to find more gadgets. The write system call takes three arguments: a socket, a buffer and a length. Arguments are passed in rdi, rsi and rdx registers, and the system call number is stored in the rax register. The following gadgets are therefore needed: 1) pop rdi; ret ( socket ) 2) pop rsi; ret ( buffer ) 3) pop rdx; ret ( length ) 4) pop rax; ret (write syscall number) 5) syscall The first optimization is the BROP gadget. Misaligned parses of it yield a pop rdi and pop rsi. So by finding a single gadget, we find two gadgets that control the first two arguments of a call. The second optimization is finding a call write. Instead of finding two gadgets (pop rax; ret and syscall) we can find a single call write instruction The problem is now reduced to finding the BROP gadget, write’s entry in the PLT and a way to control rdx for the length of the write. Any greater than zero length for write will do as one can leak the binary in multiple small chunks by chaining writes. We note that at the time of exploit, rdx may have a sane (greater than zero) value so having to control rdx may be unnecessary, but for the general case it is. Unfortunately pop rdx; ret gadgets are rare, so an optimization is to find a call to strcmp instead (again in the PLT) which sets rdx to the length of the string being compared. The optimized attack therefore requires: 1) Finding the BROP gadget. 2) Finding the PLT. Finding the entry for write. Finding the entry for strcmp. 8.2 Finding gadgets and the stop gadget The basic idea in finding gadgets remotely is to scan the application’s text segment by overwriting the saved return address with an address pointing to text and inspecting program behavior. A starting address can be found from the initial stack reading phase or 0x400000 can be used on default non-PIE Linux. Generally speaking two things will occur: the program will crash or it will hang, and in turn the connection will close or stay open. Most of the time the program will crash, but when it does not, a gadget is found.
• 13.
  13 A problem withusing this technique naively for finding gadgets is that even if the return address is overwritten with the address of a useful gadget like pop rdi; ret, the application will still likely crash because it will eventually attempt to return to the next word on the stack, likely an invalid address. The crash would cause us to discard the gadget classifying it as uninteresting. A stop gadget is anything that would cause the program to block, like an infinite loop or a blocking system call (like sleep). To scan for useful gadgets, one places the address being probed in the return address,followed by a number of stop gadgets. The next section differentiate between pop rdi; ret and pop rsi; ret. Stop gadgets need not necessarily be gadgets that “stop” the program. They are merely a signaling mechanism. Another scenario is one in which a stop gadget is already present in the stack frame. The stack will indeed already have multiple return addresses in it, and one of them may act as a stop gadget. This particular optimization is useful for preventing situations where worker processes are limited and infinite loop-type stop gadgets cause all workers to become stuck, making it possible to continue the attack (as is the case with nginx). 8.3 Identifying gadgets We now discuss how to classify gadgets. This can be done by controlling the stack layout and inspecting program behavior. We define three values that the attacker can place on the stack: Probe The address of the gadget being scanned. Stop The address of a stop gadget that will not crash. Trap The address of non-executable memory that will cause a crash (e.g., 0x0). The idea is that by varying the position of the stop and trap on the stack, one can deduce the instructions being executed by the gadget, either because the trap or stop will execute, causing a crash or no crash respectively. Here are some examples and possible stack layouts: • probe, stop, traps (trap, trap, ...). Will find gadgets that do not pop the stack like ret or xor rax, rax; ret. • probe, trap, stop, traps. Will find gadgets that pop exactly one stack word like pop rax; ret or pop rdi; ret. Figure 10 shows an illustration of this. • probe, stop, stop, stop, stop, stop, stop, stop, traps. Will find gadgets that pop up to six words (e.g., the BROP gadget). The BROP gadget has a very unique signature. It pops six items from the stack and landing in other parts of it pops fewer items from the stack so one can verify a candidate by laying out traps and stop gadgets in different combinations and checking behavior. A misaligned parse in the middle yields a pop rsp which will cause a crash and can be used to verify the gadget and further eliminate false positives. The gadget is 11 bytes long so one can skip up to 7 bytes when scanning the .text segment to find it more efficiently, landing somewhere in the middle of it. After the BROP gadget is found, the attacker can control the first two arguments (rdi and rsi) to any call.
• 14.
  14 8.4 Finding theProcedure Linking Table To control the third argument (rdx) one needs to find a call to strcmp, which sets rdx to the length of the string compared. The PLT is a jump table at the beginning of the executable used for all external calls (e.g., libc). For example, a call to strcmp will actually be a call to the PLT. The PLT will then dereference the Global Offset Table (GOT) and jump to the address stored in it. The GOT will be populated by the dynamic loader with the addresses of library calls depending on where in memory the library got loaded. The GOT is populated lazily, so the first time each PLT entry is called, it will take a slow path via dlresolve to resolve the symbol location and populate the GOT entry for the next time. The structure of the PLT is shown in Figure 11. It has a very unique signature: each entry is 16 bytes apart (16 bytes aligned) and the slow path for each entry can be run at an offset of 6 bytes. The PLT can therefore be found by scanning from the program’s origin (0x400000) or backwards from the address leaked through stack reading if the PIE flag was used. The stack layout to find a PLT entry will be: probe, stop, trap. The PLT can then be verified by seeing if neighboring entries do not crash, and if offsets of six bytes still do not cause a crash 8.5 Controlling rdx via strcmp Once the attacker finds the PLT, the question is what function calls do various entries correspond to? One of them will be strcmp, unless the program or one of its libraries does not use that function in which case the attacker can perform a “first principles” attack described later to find pop rdx; ret. The attacker can identify PLT entries by exercising each entry with different arguments and seeing how the function performs. The first two arguments can be controlled thanks to the BROP gadget. The attacker finds strcmp by finding an entry that responds to the previously mentioned signature. The PLT can be scanned in two ways. The attacker will have the address of a valid PLT entry, found previously. Once strcmp is found, the attacker can set rdx to a non-zero value by just supplying a pointer to either a PLT entry (non-zero code sequence) or the start of the ELF header (0x400000) which has seven non-zero bytes. False positives can be found when searching for strcmp. Interestingly though, in all our tests we found that strncmp and strcasecmp were found instead, and all had the effect of setting rdx to a value greater than zero, thereby fulfilling the same purpose. 8.6 Finding write The attacker so far can control the first three arguments to any call: the first two via the BROP gadget, and the third one indirectly via strcmp. There are two approaches: chaining multiple writes each with different file descriptor numbers in a single ROP chain, or opening
• 15.
  15 multiple connections andusing a relatively high file descriptor number in hope that it will match one of the connections. We use both techniques in combination. 8.7 Concluding the attack At this point the attacker can write the entire .text segment from memory to the attacker’s socket, disassemble it, and find more gadgets. The attacker can also dump the symbol table and find useful functions in the PLT like dup2 and execve. Generally speaking the attacker will need to: 1) Redirect the socket to standard input / output. The attacker can use dup2 or close, followed by either dup or fcntl(F_DUPFD). These are often in the PLT. 2) Find “/bin/sh” in memory. An effective technique is to find a writable memory region like the environment, environ, from the symbol table, and read “/bin/sh” from the attacker’s socket to that address. 3) execve the shell. If execve is not in the PLT, the attacker will need to transfer more of the binary to find a pop rax; ret and syscall gadget. 8.8 Attack summary The optimized BROP attack is as follows: 4) Find where the executable is loaded. Either 0x400000 for non-PIE executables (default) or stack read a saved return address. 5) Find a stop gadget. This is typically a blocking system call (like sleep or read) in the PLT. The attacker finds the PLT in this step too. 6) Find the BROP gadget. The attacker can now control the first two arguments to calls. 7) Find strcmp in the PLT. The attacker can now control the first three arguments to calls. 8) Find write in the PLT. The attacker can now dump the entire binary to find more gadgets. 9) Build a shellcode and exploit the server. The attack requires a single (partial) scan of the executable. The PLT is the first part of an executable and is also the first item the attacker needs to bootstrap ( i.e., finding a stop gadget). The attack complexity is therefore based on the density of BROP gadgets and how long it takes to find the PLT. 8.9 First principles attack The attack outline is: 1) Find all pop x; ret gadgets. 2) Find a syscall gadget. 3) Identify the pop gadgets previously found. The attacker starts by finding a stop gadget and all pop x; ret instructions. The difficulty is now in identifying the pop instructions and finding a syscall gadget. One must have a priori identified pop rax; ret. The solution is to chain all pop instructions found by the attacker, popping the desired system call number, and one of them will likely be rax.
• 16.
  16 At this pointthe attacker has the address of a syscall gadget and a pop rax; ret gadget. The attacker also holds a list of unidentified pops. These are identified by using the following system calls: 1) First argument (pop rdi): nanosleep(len, rem). This will cause a sleep of len nanoseconds (no crash). rem is populated if the sleep is interrupted, and it can be an invalid address as it is checked only after the sleep. 2) Second argument (pop rsi): kill(pid, sig). If sig is zero, no signal is sent, otherwise one is sent (causing a crash). 3) Third argument (pop rdx): clock_nanosleep(clock, flags, len, rem). Similar to nanosleep but takes two additional arguments, making the third argument control the sleep length. A significant optimization is that all pop rax; ret gadgets we found were misaligned parses of add rsp, 0 x58; ret.
• 17.
  17 Chapter 9: IMPLEMENTATION We implementedthe BROP attack in a tool called “Braille” that automatically goes from a crash to a remote shell. It is written in 2,000 lines of Ruby code. Braille is essentially a meta-exploit that takes a driver function that can crash a server, and figures out all the information needed to build an exploit. The interface is as simple as: try_exp(data) -> CRASH, NO_CRASH, INF The driver needs to implement the try_exp() function and guarantee that when given enough “data” bytes it will end up overwriting a saved return address on the stack. The function returns CRASH if the socket closes after sending the data, NO CRASH if the application appears to behave normally, or INF if the socket stays open for longer than a timeout. The timeout is automatically determined by the framework based on how quickly a crash is detected. The NO CRASH def try_exp(data) s = TCPSocket.new($victim, 80) req = "GET / HTTP/1.1rn" req << "Host: site.comrn" req << "Transfer-Encoding: Chunkedrn" req << "Connection: Keep-Alivern" req << "rn" req << "#{0xdeadbeefdead.to_s(16)}rn" req << "#{data}" s.write(req) if s.read() == nil return RC_CRASH else return RC_NOCRASH end end Figure 9.1 . nginx exploit driver code. It merely wraps the input string provided by Braille into an HTTP chunked request and returns a CRASH signal if the connection closes. return code is useful for probing the length of the buffer being overflowed or when stack reading. This return code is expected when no overflow occurs or when the same value was overwritten on the stack. The INF return code is expected when looking for gadgets as it indicates that the stop gadget ran (“infinite loop”). The driver code can be as simple as opening a socket to the server, and sending the data over the socket, raw. Figure 9.1 shows the code of a basic version of the nginx driver. We also implemented a generic IP fragmentation router which is useful in cases where a single large TCP read is needed to overflow a buffer.
• 18.
  18 Sending TCP segmentsout of order would be an alternative approach which may work and be more robust to firewalls. ClientHello ch; for (uint16 i = 0; i < ch.suite_len_; i += 3) { input.read(&ch.cipher_suites_[j], SUITE_LEN); j += SUITE_LEN; } input.read(ch.session_id_, ch.id_len_); if (randomLen < RAN_LEN) memset(ch.random_, 0, RAN_LEN - randomLen); input.read(&ch.random_[RAN_LEN - randomLen], randomLen);
• 19.
  19 Chapter 10: EVALUATION We testedthe BROP attack in three scenarios: 1) An open-source SSL library with a known stack vulnerability (yaSSL). This mimics the scenario where one is attacking a proprietary service that is believed to use a vulnerable open-source component. As a sample target we used an older version of MySQL that used yaSSL. 2) An open-source software with a known stack vulnerability (nginx), manually compiled from source. In this scenario the attacker knows the source of the entire server but does not hold the binary. 3) A toy closed-binary proprietary service with a stack vulnerability. This was written by a colleague and both the binary and source were kept secret. Ideally we would have tested this against a real-world proprietary service but it would have been difficult to do so legally. The vulnerabilities were as follows. 1) yaSSL: The SSL hello packet has variable length values for the cipher-suite, session id, and client random. The lengths are specified in the packet, and yaSSL copies these values into fixed sized buffers on the stack. This makes it possible to exploit the program in three ways by overflowing any of the buffers. Interestingly, the copy for the client random contains pointer arithmetic which makes it possible to write on the stack starting from the saved return address instead of being forced to overwrite the canary.. 2) nginx: HTTP requests can be sent as chunks, where each chunk has a length, followed by the chunk data. If a large chunk value is supplied, it will be stored in a signed variable typedef struct { ... off_t content_length_n; } ngx_http_headers_in_t; u_char buffer[NGX_HTTP_DISCARD_BUFFER_SIZE]; size = (size_t) ngx_min( r->headers_in.content_length_n, NGX_HTTP_DISCARD_BUFFER_SIZE); n = r->connection->recv(r->connection, buffer, size); Figure 10.1. Vulnerable code in nginx. The attacker controls content_length_n (signed) and can supply a negative value. Size will be unsigned, resulting in a large number if content_length_n is negative. buffer is 4096 bytes long, on the stack. resulting in a negative number. A check is performed if the size (now negative) is smaller than the buffer size. The size is then cast to an unsigned value passed as the length parameter
• 20.
  20 to read, makingit possible to read a large chunk into a 4096byte buffer on the stack. Figure 14 shows the vulnerable code in nginx. c) Proprietary service: When sending a short string, “OK” was read from the server. On a very long string, the connection closed. We ran Braille against all three attack scenarios, without any application-specific optimizations, and the attack succeeded in all cases. We evaluate the following aspects: 1) Performance: number of requests and time. 2) Stability: how robust the attack is. 3) Attack paired with source-code knowledge: whether having access to the source code (but not the binary) can make the attack better. A. Performance After the BROP gadget is found, finding write takes only a few additional requests, and can usually be found in approximately 2,000 requests total. At this point the attack is almost complete. One may choose to manually write very specific parts of the binary to minimize the number of requests based on the information learned. Otherwise, our Braille tool starts dumping the binary from its start, until the entire symbol table is dumped so that a shellcode can be constructed. In the case of yaSSL, it took many more requests to dump the binary because the buffer being overflowed was very short and so Braille was limited in how long the ROP chain could be. The attack clearly is noisy but we argue that if it executes fast enough, the attacker may be able to perform whatever activity he needs to do before getting caught. B. Stability MySQL is (typically) single process, multi-threaded. On a crash, a script (mysqld safe) reexecutes the server. The BROP attack works under the default configuration (no PIE, but canaries) despite the re-execution because the canary is never hit thanks to how the bug is being exercised. If compiled with the PIE flag, the attack would not work as one couldn’t read a (changing) return address from the stack to defeat ASLR. nginx has multiple worker processes and has a singlethreaded, event-based architecture. Most distributions configure four worker processes by default. This makes it a tricky scenario because an infinite loop based stop gadget would hog the worker completely, and one gets only four shots by default. The stop gadget here was returning to a higher stack frame, which avoided any blocking. C. Source-code aid If the attacker has some knowledge of the server (e.g., source code) the number of requests may be brought down. In the case of nginx, the overflow happens right after a read call so rdi (the first argument) has the socket number. The attacker can exploit this fact by calling write without setting rdi or calling dup2 to copy the file descriptor to a fixed number; by comparison the generic attack must brute-force the file descriptor number. Knowing that reads are non-blocking is also useful so that the IP fragmentation router is used to trigger the overflow.
• 21.
  21 Source knowledge helpedin the yaSSL case too. The bug could be exploited in three ways, but only one allowed circumventing canaries. Based on source information, an attacker may determine whether rdx has a sane value at the time of exploit. If so, the attacker may skip having to find strcmp to control rdx and proceed with a more optimal attack. Stack reading is a very useful tool when hacking blind. It reveals whether canaries are used, frame pointers are enabled, and possibly more. This helps in fingerprinting the distribution (based on defaults) or the environment being hacked. LIMITATIONS The BROP attack has its limitations. We applied it only to simple stack overflows. While it is a good starting point, many vulnerabilities are more complex and heap-based. Stack reading assumes that the attacker can overflow at a byte granularity and controls the last byte being overflown (e.g., a zero is not appended by the server).
• 22.
  22 Chapter 11: BROP PREVENTION Rerandomization Themost basicprotectionagainstthe BROPattack is to rerandomize canariesandASLRasoftenas possible.Theseprotectionmechanismsare effective,butserverdevelopersundermine themby not rerandomizingwhenworkerprocessescrash.The simplestmethodistoforkandexecthe processon a crash or spawn,whichrerandomizesboththe canaryandASLR. It isimportantthat anychild processesforkedare randomizedindependentlysothatany informationlearnedfromone child cannot be usedagainstanotherone. A. Sleep on crash Systemslike NetBSD’ssegvguard andgrsec’s deter_bruteforce forLinux [21] propose delayinga fork aftera segmentationfault.Thistechniquecanslow downattackssuch thatan administratorcan notice somethingiswrongandaddressthe problem. A. CompilerTechniques Many moderncompilerssupportinsertingruntime boundschecksonbuffers.Thispreventsaclass of bugs,unlike canariesthatdetectcorruptionafterthe fact or notat all if a successfullyhacker brute-forcesthe secretvalue.LLVMcontains AddressSanitizer,atool for boundschecks,anduse afterfree bugs.
• 23.
  23 Chapter 12: RELATED WORK Priorwork exists on scanning for a single ROP gadget. Goodspeed’s half-blind attack against microcontrollers relies on knowledge of (common) bootloader code, and scanning for a single gadget in an unknown portion of memory to construct an attack used for firmware extraction. The BROP attack is more generic as it is fully blind and presents techniques to find and chain multiple, different, gadgets. There has been work on attacking instruction set randomization that uses similar techniques to the BROP attack. In that work, signaling for correct/incorrect guesses was based on whether an application crashes or not. Stack reading to determine canaries is a well known technique . We generalize the stack reading technique to reading off the saved return address and frame pointer to break 64- bit ASLR. Many have noted that exploits today are multistage and require leaking information. Kingcope’s 32-bit nginx exploit for example brute-forces the location of write in the PLT to leak the binary to find gadgets and construct a binaryagnostic exploit . Figures to discuss: BROP in different OSes Windows lacks a fork-like API (it has only CreateProcess) so canaries and the text segment’s base address are guaranteed to be rerandomized after a crash,making the system more robust against BROP-like attacks. The Windows ABI also passes arguments in scratch registers (e.g., rcx,rdx) making pop gadgets for them harder to find. ASLR implementations vary by OS. Windows 8.1 and Mac OS X randomize everything by default. Unfortunately, both systems rerandomize system libraries only at reboot time. This can create a BROP-like environment for leaking pointers to system libraries. Reboots can be rare,especially on clients and laptop systems where users may suspend and resume more often than reboot. Mac OS X also only supports 16-bits of entropy for ASLR, placing it far behind other 64-bit operating systems. On Linux, the effectiveness of ASLR depends on the distribution and its PIE configuration. For
• 24.
  24 example, Ubuntu doesnot enable PIE by default, but has enabled it on a perapplication basis based on risk . BROP beyond stack overflows The BROP attack focuses on stack attacks,the simplest possible scenario. We did not consider heap exploits, though these might be possible. The bootstrap, for example, would be different, as a stack pivot gadget would have to be found after the stop gadget. A useful stack pivot gadget would be, for example, mov rax, rsp; ret assuming that rax points to an attacker-controlled buffer at the time of exploit. The attacker can now set up a stack and ROP chains in that buffer. Variance in binaries Worse for the attacker,a system may be manually compiled by the end user,making it impossible for the attacker to build a ROP chain offline as the binary is unknown. In such cases BROP is a necessity. Even if a server uses a precompiled binary, it can be difficult to determine which particular one is being used: remote OS fingerprinting reveals an approximate kernel version, not a distribution. BROP in fact can be used to fingerprint distributions and applications (e.g., based on whether canaries are present, vsyscall behavior, etc.). Remote fuzz testing The BROP attack could be a powerful tool for hacking proprietary closed-binary services when coupled with a remote fuzz tester. We note that in two of the example applications we targeted, the overflow occurred because a length was specified in the packet but a larger value was sent. It certainly is possible to write a fuzz tester that has knowledge about a protocol and attempts to overflow by supplying incorrect lengths .
• 25.
  25 Chapter 13: CONCLUSION We showthat, under the right conditions, it is possible to write exploits without any knowledge of the target binary or source code. This works for stack vulnerabilities where the server process restarts after a crash. Our attack is able to defeat ASLR, NX and stack canaries on modern 64-bit Linux servers. We present two new techniques: generalized stack reading, which defeats full ASLR on 64-bit systems, and the BROP attack, which is able to remotely find ROP gadgets. Our fully automated tool, Braille, can take under 4,000 requests to spawn a shell, under 20 minutes, tested against real versions of yaSSL+MySQL and nginx with known vulnerabilities, and a toy proprietary service running an unknown binary. We show that design patterns like forking servers with multiple worker processes can be at odds with ASLR, and that ASLR is only effective when it is applied to all code segments in the binary (including PIE). Moreover, security through obscurity, where the binary is unknown or randomized, can only slow but not prevent buffer overflow attacks.
• 26.
  26 REFERENCES [1] A. One,“Smashing The Stack For Fun And Profit,” Phrack, vol. 7, no. 49, Nov. 1996. http://phrack.com/issues.html?issue=49&id=14#article [2] IEEE paper. [3] Online: www.Quora.com [4] P. Bakke, S. Beattie, A. Grier, P. Wagle, and Q. Zhang, “Stackguard: automatic adaptive detection and prevention of buffer-overflow attacks,” in Proceedings of the 7th conference on USENIX Security Symposium - Volume 7, [5] Peach fuzzer. [Online]. Available: http://peachfuzzer.com/