The document outlines various attack vectors against the Windows kernel, including directly from user mode, through public and undocumented APIs, and by exploiting architectural flaws or bugs in kernel code. It discusses tools for static and dynamic analysis, such as debuggers, driver verifiers, and disassemblers. Examples are provided of past vulnerabilities involving buffer overflows, integer overflows, and trusting user input. Defensive measures like parameter validation, code signing, and moving functionality to user mode are assessed. Future work exploring fuzzing, virtualizing the kernel, and automated binary analysis is proposed.
The document outlines various attack vectors against the Windows kernel, including directly from user mode, through public and undocumented APIs, and by exploiting architectural flaws or bugs in kernel code. It discusses tools for static and dynamic analysis, such as debuggers, driver verifiers, and disassemblers. Examples are provided of past vulnerabilities involving buffer overflows, integer overflows, and trusting user input. Defensive measures like parameter validation, code signing, and moving functionality to user mode are assessed. Future work exploring fuzzing, virtualizing the kernel, and automated binary analysis is proposed.
Introduction Limited to Windows, and aimed at IA32: • Outline of protected mode and the kernel • Attack vectors • Useful tools • Examples • Defensive measures • Future directions Architecture Overview A long time ago in a galaxy far, far away… The progression from Intel’s 8088 to 80386, via the 80286, added: • Page and segment level protection • Call, interrupt and task gates • Privileged and sensitive instructions • Four privilege levels underlying the protection mechanisms above • 32bit support The supervisor The NT kernel provides: • Segregation of user mode processes • Protection of the kernel from user mode • Provide services to user mode and other kernel mode code • Session management and the Windows graphics subsystem The NT kernel • System call and DeviceIoControl covered • Graphics drivers – Display driver – Miniport driver • NDIS and TDI • Port objects • Windows Driver Framework • Kernel mode callbacks • Hardware interfaces – Talking to hardware – Listening to hardware A plan of attack • Directly from user mode? – CPU bugs – Operating system design • Public APIs – StartService, DeviceIoControl, ExtEscape • Undocumented APIs – ZwSystemDebugControl, ZwSetSystemInformation • Architectural flaws • Bugs in code • Subverting operating system initialization • Modifying kernel modules on disk – Viruses – DLL (export driver) injection Tools of the trade Two different approaches • Dynamic analysis – Will not guarantee results – Fuzzing awkward to automate • Static analysis – Can be complicated and time consuming – Source code very helpful • Best results achieved by combining both Static analysis • Static driver verifier • PREFast • Disassembler • Windows Driver Kit – Documentation and header files Dynamic analysis • WinDbg • Driver verifier • Miscellaneous – WinObj – NtDispatchPoints – Rootkit Hook Analyzer Getting our hands dirty I have the tools, now what? • Poor access control • Trusting user supplied data – Pointers and lengths • Typical coding bugs – Boundary conditions – Off-by-one errors • Design flaws – Expose kernel functionality or data Reverse engineering • Knowing the correct entry points means code coverage can be guaranteed • Subtle bugs are easier to find - signedness • Memory overwrites are very easy to find • Highlight areas of code more suited to fuzzing • No need to analyze a crash dump • Lack of symbolic information may prove awkward CDFS DispatchDeviceControl Source code analysis • Access to source is not common • Source code and a suitable IDE will greatly improve auditing speed • Assumptions made by the coder may help hide subtle bugs • Tools are available to help speed up the process even further • grep FIXME –r *.* CDFS DispatchDeviceControl Getting a foot in the door Kernel targets we are interested in: • Static or object function pointers • Kernel variables - MmUserProbeAddress • Descriptor tables • Return address • Code from a kernel module • I/O access map from TSS • Kernel structures – process token, loaded module list, privilege LUIDs Real world examples NT kernel compression support • Kernel runtime library exports functions to support compression – Used by SMB and NTFS • Support routines take a parameter indicating what algorithm to use – Used as an index into a function table • The table only has 8 entries, whereas the maximum index allowed is 15 – We can treat code or data as a function pointer, potentially to a user mode address Trusting user input • The following code takes a pointer from a buffer supplied by the user and trusts it – Either a sign-extended kernel stack address or an internal handle will be written there • This can be used to overwrite other code or data, allowing arbitrary code execution • User supplied pointers into: – user mode should be validated – kernel mode should be opaque, e.g. a handle An architectural flaw • A function designed to allow the modification of arbitrary memory • Exposed to unprivileged users • Provided the internal data structure can be figured out, it is then easy to exploit • Either access control to the driver, or a different architecture is needed Defensive measures Current architecture • Parameter validation • Code signing – quality control? • PatchGuard • Moving functionality into user mode – UMDF, display drivers in Vista • Restricting access to APIs – User restrictions – Privilege restrictions – Process restrictions Alternative approaches • Hypervisor – Designed to help virtualization – Provides a layer beneath the supervisor – It could be used to provide a microkernel architecture • Microkernel – Does not require virtualization hardware – Minimizes the attack surface provided by the kernel – Increases flexibility with respect to service implementation – Microsoft’s Singularity microkernel is strongly typed and uses software based protection Future work Fuzzing • Application fuzzing unlikely to crash the OS • We need to automate crash recovery and analysis: – Run in a VM, but what about real hardware? – Have bugcheck callbacks – Modify the kernel itself • Fuzzing interfaces is greatly aided by some form of static analysis Virtualizing the kernel • Provide a user mode environment that looks the same as the kernel • Implement user mode compatible APIs where necessary • Provide basic I/O, PnP, Process Support and executive functionality • Trap and handle protected and privileged code execution • Add instrumentation for analysis and logging Automated binary analysis • Model basic CPU functionality – Instead of processing a specific value, instructions work on a defined range – Instructions can modify the range stored in a register • Allows all code paths to be assessed – Large state space • Determine ranges of values that will hit certain pieces of code • Heuristic bug detection In conclusion … Summary • Current NT kernel architecture increases the likelihood of security issues • Debatable how much effort has gone into securing kernel code • Some areas of the kernel have not received much attention • There is plenty of scope for further research and tool development Questions?